int main(int argc, char** argv) { ccv_enable_default_cache(); assert(argc == 2); FILE *r = fopen(argv[1], "r"); char* file = (char*)malloc(1024); ccv_array_t* categorizeds = ccv_array_new(sizeof(ccv_categorized_t), 64, 0); size_t len = 1024; ssize_t read; while ((read = getline(&file, &len, r)) != -1) { while(read > 1 && isspace(file[read - 1])) read--; file[read] = 0; ccv_file_info_t input; input.filename = (char*)ccmalloc(1024); strncpy(input.filename, file, 1024); ccv_categorized_t categorized = ccv_categorized(0, 0, &input); ccv_array_push(categorizeds, &categorized); } fclose(r); free(file); /* MattNet parameters */ ccv_convnet_layer_param_t params[13] = { // first layer (convolutional => max pool => rnorm) { .type = CCV_CONVNET_CONVOLUTIONAL, .bias = 0, .glorot = sqrtf(2), .input = { .matrix = { .rows = 225, .cols = 225, .channels = 3, .partition = 1, }, }, .output = { .convolutional = { .count = 96, .strides = 2, .border = 1, .rows = 7, .cols = 7, .channels = 3, .partition = 2, }, }, }, { .type = CCV_CONVNET_LOCAL_RESPONSE_NORM,
ccv_nnc_graph_exec_symbol_t ccv_nnc_symbolic_graph_while(ccv_nnc_symbolic_graph_t* const graph, ccv_nnc_symbolic_graph_t* const while_graph, const char* const name) { assert(while_graph->p == 0); assert(while_graph->p_idx == 0); ccv_nnc_cmd_t cmd = ccv_nnc_cmd(CCV_NNC_GRAPH_FORWARD, 0, CMD_GENERIC(), 0); // Added one more symbol. ccv_nnc_graph_exec_symbol_t symbol = ccv_nnc_graph_exec_symbol_new(graph, cmd, 0, 0, 0, 0, name); // Assigning graph_ref to it. if (!graph->sub_graphs) graph->sub_graphs = ccv_array_new(sizeof(ccv_nnc_symbolic_graph_t*), 1, 0); ccv_array_push(graph->sub_graphs, &while_graph); ccv_nnc_graph_exec_symbol_info_t* symbol_info = (ccv_nnc_graph_exec_symbol_info_t*)ccv_array_get(graph->exec_symbol_info, symbol.d); // Note the extra allocation (the ccv_array_t only holds a pointer to ccv_nnc_symbolic_graph_t*). // In this way, we can get the while graph and don't have to worry about it will be an invalid pointer once // the array expands (another while graph allocated). symbol_info->graph_ref = graph->sub_graphs->rnum; while_graph->p_idx = graph->sub_graphs->rnum; while_graph->exec_idx = symbol.d + 1; while_graph->p = graph; return symbol; }
int main(int argc, char** argv) { static struct option image_net_options[] = { /* help */ {"help", 0, 0, 0}, /* required parameters */ {"train-list", 1, 0, 0}, {"test-list", 1, 0, 0}, {"working-dir", 1, 0, 0}, /* optional parameters */ {"base-dir", 1, 0, 0}, {"max-epoch", 1, 0, 0}, {"iterations", 1, 0, 0}, {0, 0, 0, 0} }; char* train_list = 0; char* test_list = 0; char* working_dir = 0; char* base_dir = 0; ccv_convnet_train_param_t train_params = { .max_epoch = 100, .mini_batch = 64, .sgd_frequency = 1, // do sgd every sgd_frequency batches (mini_batch * device_count * sgd_frequency) .iterations = 50000, .device_count = 4, .peer_access = 1, .symmetric = 1, .image_manipulation = 0.2, .color_gain = 0.001, .input = { .min_dim = 257, .max_dim = 257, }, }; int i, c; while (getopt_long_only(argc, argv, "", image_net_options, &c) != -1) { switch (c) { case 0: exit_with_help(); case 1: train_list = optarg; break; case 2: test_list = optarg; break; case 3: working_dir = optarg; break; case 4: base_dir = optarg; break; case 5: train_params.max_epoch = atoi(optarg); break; case 6: train_params.iterations = atoi(optarg); break; } } if (!train_list || !test_list || !working_dir) exit_with_help(); ccv_enable_default_cache(); FILE *r0 = fopen(train_list, "r"); assert(r0 && "train-list doesn't exists"); FILE* r1 = fopen(test_list, "r"); assert(r1 && "test-list doesn't exists"); char* file = (char*)malloc(1024); int dirlen = (base_dir != 0) ? strlen(base_dir) + 1 : 0; ccv_array_t* categorizeds = ccv_array_new(sizeof(ccv_categorized_t), 64, 0); while (fscanf(r0, "%d %s", &c, file) != EOF) { char* filename = (char*)ccmalloc(1024); if (base_dir != 0) { strncpy(filename, base_dir, 1024); filename[dirlen - 1] = '/'; } strncpy(filename + dirlen, file, 1024 - dirlen); ccv_file_info_t file_info = { .filename = filename, }; // imageNet's category class starts from 1, thus, minus 1 to get 0-index ccv_categorized_t categorized = ccv_categorized(c - 1, 0, &file_info); ccv_array_push(categorizeds, &categorized); } fclose(r0); ccv_array_t* tests = ccv_array_new(sizeof(ccv_categorized_t), 64, 0); while (fscanf(r1, "%d %s", &c, file) != EOF) { char* filename = (char*)ccmalloc(1024); if (base_dir != 0) { strncpy(filename, base_dir, 1024); filename[dirlen - 1] = '/'; } strncpy(filename + dirlen, file, 1024 - dirlen); ccv_file_info_t file_info = { .filename = filename, }; // imageNet's category class starts from 1, thus, minus 1 to get 0-index ccv_categorized_t categorized = ccv_categorized(c - 1, 0, &file_info); ccv_array_push(tests, &categorized); } fclose(r1); free(file); // #define model_params vgg_d_params #define model_params matt_c_params int depth = sizeof(model_params) / sizeof(ccv_convnet_layer_param_t); ccv_convnet_t* convnet = ccv_convnet_new(1, ccv_size(257, 257), model_params, depth); if (ccv_convnet_verify(convnet, 1000) == 0) { ccv_convnet_layer_train_param_t layer_params[depth]; memset(layer_params, 0, sizeof(layer_params)); for (i = 0; i < depth; i++) { layer_params[i].w.decay = 0.0005; layer_params[i].w.learn_rate = 0.01; layer_params[i].w.momentum = 0.9; layer_params[i].bias.decay = 0; layer_params[i].bias.learn_rate = 0.01; layer_params[i].bias.momentum = 0.9; } // set the two full connect layers to last with dropout rate at 0.5 for (i = depth - 3; i < depth - 1; i++) layer_params[i].dor = 0.5; train_params.layer_params = layer_params; ccv_set_cli_output_levels(ccv_cli_output_level_and_above(CCV_CLI_INFO)); ccv_convnet_supervised_train(convnet, categorizeds, tests, working_dir, train_params); } else { PRINT(CCV_CLI_ERROR, "Invalid convnet configuration\n"); } ccv_convnet_free(convnet); ccv_disable_cache(); return 0; }
static void _ccv_set_union_mser(ccv_dense_matrix_t* a, ccv_dense_matrix_t* h, ccv_dense_matrix_t* b, ccv_array_t* seq, ccv_mser_param_t params) { assert(params.direction == CCV_BRIGHT_TO_DARK || params.direction == CCV_DARK_TO_BRIGHT); int v, i, j; ccv_mser_node_t* node = (ccv_mser_node_t*)ccmalloc(sizeof(ccv_mser_node_t) * a->rows * a->cols); ccv_mser_node_t** rnode = (ccv_mser_node_t**)ccmalloc(sizeof(ccv_mser_node_t*) * a->rows * a->cols); if (params.range <= 0) params.range = 255; // put it in a block so that the memory allocated can be released in the end int* buck = (int*)alloca(sizeof(int) * (params.range + 2)); memset(buck, 0, sizeof(int) * (params.range + 2)); ccv_mser_node_t* pnode = node; // this for_block is the only computation that can be shared between dark to bright and bright to dark // two MSER alternatives, and it only occupies 10% of overall time, we won't share this computation // at all (also, we need to reinitialize node for the two passes anyway). if (h != 0) { unsigned char* aptr = a->data.u8; unsigned char* hptr = h->data.u8; #define for_block(_for_get_a, _for_get_h) \ for (i = 0; i < a->rows; i++) \ { \ for (j = 0; j < a->cols; j++) \ if (!_for_get_h(hptr, j, 0)) \ ++buck[_for_get_a(aptr, j, 0)]; \ aptr += a->step; \ hptr += h->step; \ } \ for (i = 1; i <= params.range; i++) \ buck[i] += buck[i - 1]; \ buck[params.range + 1] = buck[params.range]; \ aptr = a->data.u8; \ hptr = h->data.u8; \ for (i = 0; i < a->rows; i++) \ { \ for (j = 0; j < a->cols; j++) \ { \ _ccv_mser_init_node(pnode, j, i); \ if (!_for_get_h(hptr, j, 0)) \ rnode[--buck[_for_get_a(aptr, j, 0)]] = pnode; \ else \ pnode->shortcut = 0; /* this means the pnode is not available */ \ ++pnode; \ } \ aptr += a->step; \ hptr += h->step; \ } ccv_matrix_getter_integer_only_a(a->type, ccv_matrix_getter_integer_only, h->type, for_block); #undef for_block } else { unsigned char* aptr = a->data.u8; #define for_block(_, _for_get) \ for (i = 0; i < a->rows; i++) \ { \ for (j = 0; j < a->cols; j++) \ ++buck[_for_get(aptr, j, 0)]; \ aptr += a->step; \ } \ for (i = 1; i <= params.range; i++) \ buck[i] += buck[i - 1]; \ buck[params.range + 1] = buck[params.range]; \ aptr = a->data.u8; \ for (i = 0; i < a->rows; i++) \ { \ for (j = 0; j < a->cols; j++) \ { \ _ccv_mser_init_node(pnode, j, i); \ rnode[--buck[_for_get(aptr, j, 0)]] = pnode; \ ++pnode; \ } \ aptr += a->step; \ } ccv_matrix_getter_integer_only(a->type, for_block); #undef for_block } ccv_array_t* history_list = ccv_array_new(sizeof(ccv_mser_history_t), 64, 0); for (v = 0; v <= params.range; v++) { int range_segment = buck[params.direction == CCV_DARK_TO_BRIGHT ? v : params.range - v]; int range_segment_cap = buck[params.direction == CCV_DARK_TO_BRIGHT ? v + 1 : params.range - v + 1]; for (i = range_segment; i < range_segment_cap; i++) { pnode = rnode[i]; // try to merge pnode with its neighbors static int dx[] = {-1, 0, 1, -1, 1, -1, 0, 1}; static int dy[] = {-1, -1, -1, 0, 0, 1, 1, 1}; ccv_mser_node_t* node0 = _ccv_mser_find_root(pnode); for (j = 0; j < 8; j++) { int x = dx[j] + pnode->point.x; int y = dy[j] + pnode->point.y; if (x >= 0 && x < a->cols && y >= 0 && y < a->rows) { ccv_mser_node_t* nnode = pnode + dx[j] + dy[j] * a->cols; if (nnode->shortcut == 0) // this is a void node, skip continue; ccv_mser_node_t* node1 = _ccv_mser_find_root(nnode); if (node0 != node1) { // grep the extended root information ccv_mser_history_t* root0 = (node0->root >= 0) ? (ccv_mser_history_t*)ccv_array_get(history_list, node0->root) : 0; ccv_mser_history_t* root1 = (node1->root >= 0) ? (ccv_mser_history_t*)ccv_array_get(history_list, node1->root) : 0; // swap the node if root1 has higher rank, or larger in size, or root0 is non-existent if ((root0 && root1 && (root1->value > root0->value || (root1->value == root0->value && root1->rank > root0->rank) || (root1->value == root0->value && root1->rank == root0->rank && root1->size > root0->size))) || (root1 && !root0)) { ccv_mser_node_t* exnode = node0; node0 = node1; node1 = exnode; ccv_mser_history_t* root = root0; root0 = root1; root1 = root; } if (!root0) { ccv_mser_history_t root = { .rank = 0, .size = 1, .value = v, .shortcut = history_list->rnum, .parent = history_list->rnum, .head = node0, .tail = node1 }; node0->root = history_list->rnum; ccv_array_push(history_list, &root); root0 = (ccv_mser_history_t*)ccv_array_get(history_list, history_list->rnum - 1); assert(node1->root == -1); } else if (root0->value < v) { // conceal the old root as history (er), making a new one and pointing to it root0->shortcut = root0->parent = history_list->rnum; ccv_mser_history_t root = *root0; root.value = v; node0->root = history_list->rnum; ccv_array_push(history_list, &root); root0 = (ccv_mser_history_t*)ccv_array_get(history_list, history_list->rnum - 1); root1 = (node1->root >= 0) ? (ccv_mser_history_t*)ccv_array_get(history_list, node1->root) : 0; // the memory may be reallocated root0->rank = ccv_max(root0->rank, (root1 ? root1->rank : 0)) + 1; } if (root1) { if (root1->value < root0->value) // in this case, root1 is sealed as well root1->parent = node0->root; // thus, if root1->parent == itself && root1->shortcut != itself // it is voided, and not sealed root1->shortcut = node0->root; } // merge the two node1->shortcut = node0; root0->size += root1 ? root1->size : 1; /* insert one endless double link list to another, see illustration: * 0->1->2->3->4->5->0 * a->b->c->d->a * set 5.next (0.prev.next) point to a * set 0.prev point to d * set d.next (a.prev.next) point to 0 * set a.prev point to 5 * the result endless double link list will be: * 0->1->2->3->4->5->a->b->c->d->0 */ node0->prev->next = node1; ccv_mser_node_t* prev = node0->prev; node0->prev = node1->prev; node1->prev->next = node0; // consider self-referencing node1->prev = prev; root0->head = node0; root0->tail = node0->prev; } } }
int main(int argc, char** argv) { static struct option scd_options[] = { /* help */ {"help", 0, 0, 0}, /* required parameters */ {"positive-list", 1, 0, 0}, {"background-list", 1, 0, 0}, {"negative-count", 1, 0, 0}, {"working-dir", 1, 0, 0}, /* optional parameters */ {0, 0, 0, 0} }; char* positive_list = 0; char* background_list = 0; char* working_dir = 0; char* base_dir = 0; int negative_count = 0; int k; while (getopt_long_only(argc, argv, "", scd_options, &k) != -1) { switch (k) { case 0: exit_with_help(); case 1: positive_list = optarg; break; case 2: background_list = optarg; break; case 3: negative_count = atoi(optarg); break; case 4: working_dir = optarg; break; case 5: base_dir = optarg; } } assert(positive_list != 0); assert(background_list != 0); assert(working_dir != 0); assert(negative_count > 0); FILE* r0 = fopen(positive_list, "r"); assert(r0 && "positive-list doesn't exists"); FILE* r1 = fopen(background_list, "r"); assert(r1 && "background-list doesn't exists"); int dirlen = (base_dir != 0) ? strlen(base_dir) + 1 : 0; ccv_array_t* posfiles = ccv_array_new(sizeof(ccv_file_info_t), 32, 0); char* file = (char*)malloc(1024); size_t len = 1024; ssize_t read; while ((read = getline(&file, &len, r0)) != -1) { while(read > 1 && isspace(file[read - 1])) read--; file[read] = 0; ccv_file_info_t file_info; file_info.filename = (char*)ccmalloc(1024); if (base_dir != 0) { strncpy(file_info.filename, base_dir, 1024); file_info.filename[dirlen - 1] = '/'; } strncpy(file_info.filename + dirlen, file, 1024 - dirlen); ccv_array_push(posfiles, &file_info); } fclose(r0); ccv_array_t* hard_mine = (ccv_array_t*)ccv_array_new(sizeof(ccv_file_info_t), 32, 0); while ((read = getline(&file, &len, r1)) != -1) { while(read > 1 && isspace(file[read - 1])) read--; file[read] = 0; ccv_file_info_t file_info; file_info.filename = (char*)ccmalloc(1024); if (base_dir != 0) { strncpy(file_info.filename, base_dir, 1024); file_info.filename[dirlen - 1] = '/'; } strncpy(file_info.filename + dirlen, file, 1024 - dirlen); ccv_array_push(hard_mine, &file_info); } fclose(r1); free(file); ccv_scd_train_param_t params = { .boosting = 10, .size = ccv_size(48, 48), .feature = { .base = ccv_size(8, 8), .range_through = 4, .step_through = 4, }, .stop_criteria = { .hit_rate = 0.995, .false_positive_rate = 0.5, .accu_false_positive_rate = 1e-7, .auc_crit = 1e-5, .maximum_feature = 2048, .prune_stage = 3, .prune_feature = 4, }, .weight_trimming = 0.98,
int main(int argc, char** argv) { static struct option icf_options[] = { /* help */ {"help", 0, 0, 0}, /* required parameters */ {"positive-list", 1, 0, 0}, {"background-list", 1, 0, 0}, {"validate-list", 1, 0, 0}, {"working-dir", 1, 0, 0}, {"negative-count", 1, 0, 0}, {"positive-count", 1, 0, 0}, {"acceptance", 1, 0, 0}, {"size", 1, 0, 0}, {"feature-size", 1, 0, 0}, {"weak-classifier-count", 1, 0, 0}, /* optional parameters */ {"base-dir", 1, 0, 0}, {"grayscale", 1, 0, 0}, {"margin", 1, 0, 0}, {"deform-shift", 1, 0, 0}, {"deform-angle", 1, 0, 0}, {"deform-scale", 1, 0, 0}, {"min-dimension", 1, 0, 0}, {"bootstrap", 1, 0, 0}, {0, 0, 0, 0} }; char* positive_list = 0; char* background_list = 0; char* validate_list = 0; char* working_dir = 0; char* base_dir = 0; int negative_count = 0; int positive_count = 0; ccv_icf_new_param_t params = { .grayscale = 0, .margin = ccv_margin(0, 0, 0, 0), .size = ccv_size(0, 0), .deform_shift = 1, .deform_angle = 0, .deform_scale = 0.075, .feature_size = 0, .weak_classifier = 0, .min_dimension = 2, .bootstrap = 3, .detector = ccv_icf_default_params, }; params.detector.step_through = 4; // for faster negatives bootstrap time int i, k; char* token; char* saveptr; while (getopt_long_only(argc, argv, "", icf_options, &k) != -1) { switch (k) { case 0: exit_with_help(); case 1: positive_list = optarg; break; case 2: background_list = optarg; break; case 3: validate_list = optarg; break; case 4: working_dir = optarg; break; case 5: negative_count = atoi(optarg); break; case 6: positive_count = atoi(optarg); break; case 7: params.acceptance = atof(optarg); break; case 8: token = strtok_r(optarg, "x", &saveptr); params.size.width = atoi(token); token = strtok_r(0, "x", &saveptr); params.size.height = atoi(token); break; case 9: params.feature_size = atoi(optarg); break; case 10: params.weak_classifier = atoi(optarg); break; case 11: base_dir = optarg; break; case 12: params.grayscale = !!atoi(optarg); break; case 13: token = strtok_r(optarg, ",", &saveptr); params.margin.left = atoi(token); token = strtok_r(0, ",", &saveptr); params.margin.top = atoi(token); token = strtok_r(0, ",", &saveptr); params.margin.right = atoi(token); token = strtok_r(0, ",", &saveptr); params.margin.bottom = atoi(token); break; case 14: params.deform_shift = atof(optarg); break; case 15: params.deform_angle = atof(optarg); break; case 16: params.deform_scale = atof(optarg); break; case 17: params.min_dimension = atoi(optarg); break; case 18: params.bootstrap = atoi(optarg); break; } } assert(positive_list != 0); assert(background_list != 0); assert(validate_list != 0); assert(working_dir != 0); assert(positive_count > 0); assert(negative_count > 0); assert(params.size.width > 0); assert(params.size.height > 0); ccv_enable_cache(512 * 1024 * 1024); FILE* r0 = fopen(positive_list, "r"); assert(r0 && "positive-list doesn't exists"); FILE* r1 = fopen(background_list, "r"); assert(r1 && "background-list doesn't exists"); FILE* r2 = fopen(validate_list, "r"); assert(r2 && "validate-list doesn't exists"); char* file = (char*)malloc(1024); ccv_decimal_pose_t pose; int dirlen = (base_dir != 0) ? strlen(base_dir) + 1 : 0; ccv_array_t* posfiles = ccv_array_new(sizeof(ccv_file_info_t), 32, 0); // roll pitch yaw while (fscanf(r0, "%s %f %f %f %f %f %f %f", file, &pose.x, &pose.y, &pose.a, &pose.b, &pose.roll, &pose.pitch, &pose.yaw) != EOF) { ccv_file_info_t file_info; file_info.filename = (char*)ccmalloc(1024); if (base_dir != 0) { strncpy(file_info.filename, base_dir, 1024); file_info.filename[dirlen - 1] = '/'; } strncpy(file_info.filename + dirlen, file, 1024 - dirlen); file_info.pose = pose; ccv_array_push(posfiles, &file_info); } fclose(r0); size_t len = 1024; ssize_t read; ccv_array_t* bgfiles = (ccv_array_t*)ccv_array_new(sizeof(ccv_file_info_t), 32, 0); while ((read = getline(&file, &len, r1)) != -1) { while(read > 1 && isspace(file[read - 1])) read--; file[read] = 0; ccv_file_info_t file_info; file_info.filename = (char*)ccmalloc(1024); if (base_dir != 0) { strncpy(file_info.filename, base_dir, 1024); file_info.filename[dirlen - 1] = '/'; } strncpy(file_info.filename + dirlen, file, 1024 - dirlen); ccv_array_push(bgfiles, &file_info); } fclose(r1); ccv_array_t* validatefiles = ccv_array_new(sizeof(ccv_file_info_t), 32, 0); // roll pitch yaw while (fscanf(r2, "%s %f %f %f %f %f %f %f", file, &pose.x, &pose.y, &pose.a, &pose.b, &pose.roll, &pose.pitch, &pose.yaw) != EOF) { ccv_file_info_t file_info; file_info.filename = (char*)ccmalloc(1024); if (base_dir != 0) { strncpy(file_info.filename, base_dir, 1024); file_info.filename[dirlen - 1] = '/'; } strncpy(file_info.filename + dirlen, file, 1024 - dirlen); file_info.pose = pose; ccv_array_push(validatefiles, &file_info); } fclose(r2); free(file); ccv_icf_classifier_cascade_t* classifier = ccv_icf_classifier_cascade_new(posfiles, positive_count, bgfiles, negative_count, validatefiles, working_dir, params); char filename[1024]; snprintf(filename, 1024, "%s/final-cascade", working_dir); ccv_icf_write_classifier_cascade(classifier, filename); for (i = 0; i < posfiles->rnum; i++) { ccv_file_info_t* file_info = (ccv_file_info_t*)ccv_array_get(posfiles, i); free(file_info->filename); } ccv_array_free(posfiles); for (i = 0; i < bgfiles->rnum; i++) { ccv_file_info_t* file_info = (ccv_file_info_t*)ccv_array_get(bgfiles, i); free(file_info->filename); } ccv_array_free(bgfiles); for (i = 0; i < validatefiles->rnum; i++) { ccv_file_info_t* file_info = (ccv_file_info_t*)ccv_array_get(validatefiles, i); free(file_info->filename); } ccv_array_free(validatefiles); ccv_disable_cache(); return 0; }
ccv_nnc_tensor_symbol_t ccv_nnc_find_tensor_symbol_from_graph(const ccv_nnc_symbolic_graph_t* const graph, const ccv_nnc_tensor_symbol_t symbol) { if (symbol.graph == graph) return symbol; const ccv_nnc_symbolic_graph_t* curr_graph = symbol.graph; ccv_nnc_tensor_symbol_info_t* const symbol_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(symbol.graph->tensor_symbol_info, symbol.d); assert(symbol.d >= 0 && symbol.d < curr_graph->tensor_symbol_info->rnum); while (curr_graph && curr_graph != graph) curr_graph = curr_graph->p; if (curr_graph) { curr_graph = symbol.graph; ccv_nnc_tensor_symbol_info_t* curr_symbol_info = symbol_info; ccv_nnc_tensor_symbol_t curr_symbol = symbol; while (curr_graph != graph) { ccv_nnc_symbolic_graph_t* const p = curr_graph->p; // I need to find the symbol, it must exist. assert(curr_symbol_info->p_ref); // Move on. curr_symbol.d = curr_symbol_info->p_ref - 1; curr_symbol.graph = p; assert(curr_symbol.d >= 0 && curr_symbol.d < p->tensor_symbol_info->rnum); curr_symbol_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(p->tensor_symbol_info, curr_symbol.d); curr_graph = p; } return curr_symbol; } // Otherwise, if the symbol is in the parent graph, this is a bit more expensive because I need to keep a trace stack. curr_graph = graph; ccv_array_t* trace = ccv_array_new(sizeof(int), 0, 0); while (curr_graph && curr_graph != symbol.graph) { const int p_idx = curr_graph->p_idx - 1; ccv_array_push(trace, &p_idx); curr_graph = curr_graph->p; } // If it is not in both the parent graph and the sub-graph, the input is invalid. assert(curr_graph); curr_graph = symbol.graph; ccv_nnc_tensor_symbol_info_t* curr_symbol_info = symbol_info; ccv_nnc_tensor_symbol_t curr_symbol = symbol; // The graph is a sub graph of the symbol passed in. int i; for (i = trace->rnum - 1; i >= 0; i--) { const int p_idx = *(int*)ccv_array_get(trace, i); assert(p_idx >= 0); assert(curr_graph->sub_graphs); assert(curr_symbol_info->s_ref); assert(p_idx >= 0 && p_idx < curr_symbol_info->s_ref->rnum); const int s_idx = *(int*)ccv_array_get(curr_symbol_info->s_ref, p_idx); ccv_nnc_symbolic_graph_t* const s = *(ccv_nnc_symbolic_graph_t**)ccv_array_get(curr_graph->sub_graphs, p_idx); // I need to find the symbol, it must exist. assert(s_idx); curr_symbol.d = s_idx - 1; curr_symbol.graph = s; assert(curr_symbol.d >= 0 && curr_symbol.d < s->tensor_symbol_info->rnum); curr_symbol_info = (ccv_nnc_tensor_symbol_info_t*)ccv_array_get(s->tensor_symbol_info, curr_symbol.d); // Move on. curr_graph = s; } ccv_array_free(trace); return curr_symbol; }
static ccv_nnc_graph_t* ccv_nnc_simple_graph(ccv_convnet_t* convnet, ccv_nnc_tensor_t* input, ccv_nnc_tensor_t* output, ccv_nnc_graph_exec_t* source, ccv_nnc_graph_exec_t* dest, ccv_array_t* tensors) { int i; // We only create the graph compute to the last fc layer. ccv_nnc_graph_t* vgg = ccv_nnc_graph_new(); ccv_nnc_graph_exec_t previous_exec; for (i = 0; i < convnet->count; i++) { ccv_convnet_layer_t* layer = convnet->layers + i; int rows, cols, partition; ccv_convnet_make_output(layer, layer->input.matrix.rows, layer->input.matrix.cols, &rows, &cols, &partition); ccv_nnc_tensor_t* tensor = output; if (i < convnet->count - 1) { if (layer->type == CCV_CONVNET_FULL_CONNECT) tensor = ccv_nnc_tensor_new(0, ONE_CPU_TENSOR(rows * cols * partition), 0); else tensor = ccv_nnc_tensor_new(0, ONE_CPU_TENSOR(rows, cols, (layer->type == CCV_CONVNET_CONVOLUTIONAL ? layer->net.convolutional.count : layer->input.matrix.channels)), 0); ccv_array_push(tensors, &tensor); } ccv_nnc_graph_exec_t exec = {0}; if (layer->type == CCV_CONVNET_CONVOLUTIONAL) { ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, ONE_CPU_TENSOR(layer->net.convolutional.count, layer->net.convolutional.rows, layer->net.convolutional.cols, layer->net.convolutional.channels), 0); memcpy(w->data.f32, layer->w, layer->wnum * sizeof(float)); ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, ONE_CPU_TENSOR(layer->net.convolutional.count), 0); memcpy(bias->data.f32, layer->bias, layer->net.convolutional.count * sizeof(float)); ccv_array_push(tensors, &w); ccv_array_push(tensors, &bias); ccv_nnc_cmd_t cmd = CMD_CONVOLUTION_FORWARD(layer->net.convolutional.count, layer->net.convolutional.rows, layer->net.convolutional.cols, layer->net.convolutional.channels); ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, input->info, tensor->info); cmd = ccv_nnc_cmd_autotune(cmd, 0, hint, 0, TENSOR_LIST(input, w, bias), TENSOR_LIST(tensor), 0); exec = ccv_nnc_graph_exec_new(vgg, cmd, hint, TENSOR_LIST(input, w, bias), TENSOR_LIST(tensor)); } else if (layer->type == CCV_CONVNET_MAX_POOL) { ccv_nnc_cmd_t cmd = ccv_nnc_cmd(CCV_NNC_MAX_POOL_FORWARD, 0, CMD_GENERIC(layer->net.pool.size, layer->net.pool.size, layer->input.matrix.channels), 0); ccv_nnc_hint_t hint = ccv_nnc_hint_auto(cmd.info, input->info, tensor->info); exec = ccv_nnc_graph_exec_new(vgg, cmd, hint, TENSOR_LIST(input), TENSOR_LIST(tensor)); } else if (layer->type == CCV_CONVNET_FULL_CONNECT) { ccv_nnc_tensor_t* w = ccv_nnc_tensor_new(0, ONE_CPU_TENSOR(layer->net.full_connect.count, layer->input.node.count), 0); memcpy(w->data.f32, layer->w, layer->wnum * sizeof(float)); ccv_nnc_tensor_t* bias = ccv_nnc_tensor_new(0, ONE_CPU_TENSOR(layer->net.full_connect.count), 0); memcpy(bias->data.f32, layer->bias, layer->net.full_connect.count * sizeof(float)); ccv_array_push(tensors, &w); ccv_array_push(tensors, &bias); ccv_nnc_cmd_t cmd = ccv_nnc_cmd(CCV_NNC_GEMM_FORWARD, 0, CMD_GEMM(layer->net.full_connect.count), 0); // If the input is not what I expected (array), reshape it. if (input->info.dim[0] != ccv_nnc_tensor_count(input->info)) { input = ccv_nnc_tensor_new(input->data.u8, ONE_CPU_TENSOR(ccv_nnc_tensor_count(input->info)), 0); ccv_array_push(tensors, &input); } cmd = ccv_nnc_cmd_autotune(cmd, 0, ccv_nnc_no_hint, 0, TENSOR_LIST(input, w, bias), TENSOR_LIST(tensor), 0); exec = ccv_nnc_graph_exec_new(vgg, cmd, ccv_nnc_no_hint, TENSOR_LIST(input, w, bias), TENSOR_LIST(tensor)); } else { assert("unreachable"); } if (i != 0) ccv_nnc_graph_exec_concat(vgg, previous_exec, exec); previous_exec = exec; if (i == 0) *source = exec; if (i < convnet->count - 1 && (layer->type == CCV_CONVNET_CONVOLUTIONAL || layer->type == CCV_CONVNET_FULL_CONNECT)) { // Create the ReLU layer. ccv_nnc_cmd_param_t cmd_params = {}; ccv_nnc_cmd_t cmd = ccv_nnc_cmd(CCV_NNC_RELU_FORWARD, 0, cmd_params, 0); exec = ccv_nnc_graph_exec_new(vgg, cmd, ccv_nnc_no_hint, TENSOR_LIST(tensor), TENSOR_LIST(tensor)); ccv_nnc_graph_exec_concat(vgg, previous_exec, exec); previous_exec = exec; } if (i == convnet->count - 1) *dest = exec; // This is the input of next layer. input = tensor; } return vgg; }
ccv_array_t* ccv_bbf_detect_objects(ccv_dense_matrix_t* a, ccv_bbf_classifier_cascade_t** _cascade, int count, ccv_bbf_param_t params) { int hr = a->rows / ENDORSE(params.size.height); int wr = a->cols / ENDORSE(params.size.width); double scale = pow(2., 1. / (params.interval + 1.)); APPROX int next = params.interval + 1; int scale_upto = (int)(log((double)ccv_min(hr, wr)) / log(scale)); ccv_dense_matrix_t** pyr = (ccv_dense_matrix_t**)alloca(ENDORSE(scale_upto + next * 2) * 4 * sizeof(ccv_dense_matrix_t*)); memset(pyr, 0, (scale_upto + next * 2) * 4 * sizeof(ccv_dense_matrix_t*)); if (ENDORSE(params.size.height != _cascade[0]->size.height || params.size.width != _cascade[0]->size.width)) ccv_resample(a, &pyr[0], 0, a->rows * ENDORSE(_cascade[0]->size.height / params.size.height), a->cols * ENDORSE(_cascade[0]->size.width / params.size.width), CCV_INTER_AREA); else pyr[0] = a; APPROX int i; int j, k, t, x, y, q; for (i = 1; ENDORSE(i < ccv_min(params.interval + 1, scale_upto + next * 2)); i++) ccv_resample(pyr[0], &pyr[i * 4], 0, (int)(pyr[0]->rows / pow(scale, i)), (int)(pyr[0]->cols / pow(scale, i)), CCV_INTER_AREA); for (i = next; ENDORSE(i < scale_upto + next * 2); i++) ccv_sample_down(pyr[i * 4 - next * 4], &pyr[i * 4], 0, 0, 0); if (params.accurate) for (i = next * 2; ENDORSE(i < scale_upto + next * 2); i++) { ccv_sample_down(pyr[i * 4 - next * 4], &pyr[i * 4 + 1], 0, 1, 0); ccv_sample_down(pyr[i * 4 - next * 4], &pyr[i * 4 + 2], 0, 0, 1); ccv_sample_down(pyr[i * 4 - next * 4], &pyr[i * 4 + 3], 0, 1, 1); } ccv_array_t* idx_seq; ccv_array_t* seq = ccv_array_new(sizeof(ccv_comp_t), 64, 0); ccv_array_t* seq2 = ccv_array_new(sizeof(ccv_comp_t), 64, 0); ccv_array_t* result_seq = ccv_array_new(sizeof(ccv_comp_t), 64, 0); /* detect in multi scale */ for (t = 0; t < count; t++) { ccv_bbf_classifier_cascade_t* cascade = _cascade[t]; APPROX float scale_x = (float) params.size.width / (float) cascade->size.width; APPROX float scale_y = (float) params.size.height / (float) cascade->size.height; ccv_array_clear(seq); for (i = 0; ENDORSE(i < scale_upto); i++) { APPROX int dx[] = {0, 1, 0, 1}; APPROX int dy[] = {0, 0, 1, 1}; APPROX int i_rows = pyr[i * 4 + next * 8]->rows - ENDORSE(cascade->size.height >> 2); APPROX int steps[] = { pyr[i * 4]->step, pyr[i * 4 + next * 4]->step, pyr[i * 4 + next * 8]->step }; APPROX int i_cols = pyr[i * 4 + next * 8]->cols - ENDORSE(cascade->size.width >> 2); int paddings[] = { pyr[i * 4]->step * 4 - i_cols * 4, pyr[i * 4 + next * 4]->step * 2 - i_cols * 2, pyr[i * 4 + next * 8]->step - i_cols }; for (q = 0; q < (params.accurate ? 4 : 1); q++) { APPROX unsigned char* u8[] = { pyr[i * 4]->data.u8 + dx[q] * 2 + dy[q] * pyr[i * 4]->step * 2, pyr[i * 4 + next * 4]->data.u8 + dx[q] + dy[q] * pyr[i * 4 + next * 4]->step, pyr[i * 4 + next * 8 + q]->data.u8 }; for (y = 0; ENDORSE(y < i_rows); y++) { for (x = 0; ENDORSE(x < i_cols); x++) { APPROX float sum; APPROX int flag = 1; ccv_bbf_stage_classifier_t* classifier = cascade->stage_classifier; for (j = 0; j < ENDORSE(cascade->count); ++j, ++classifier) { sum = 0; APPROX float* alpha = classifier->alpha; ccv_bbf_feature_t* feature = classifier->feature; for (k = 0; k < ENDORSE(classifier->count); ++k, alpha += 2, ++feature) sum += alpha[_ccv_run_bbf_feature(feature, ENDORSE(steps), u8)]; if (ENDORSE(sum) < ENDORSE(classifier->threshold)) { flag = 0; break; } } if (ENDORSE(flag)) { ccv_comp_t comp; comp.rect = ccv_rect((int)((x * 4 + dx[q] * 2) * scale_x + 0.5), (int)((y * 4 + dy[q] * 2) * scale_y + 0.5), (int)(cascade->size.width * scale_x + 0.5), (int)(cascade->size.height * scale_y + 0.5)); comp.neighbors = 1; comp.classification.id = t; comp.classification.confidence = sum; ccv_array_push(seq, &comp); } u8[0] += 4; u8[1] += 2; u8[2] += 1; } u8[0] += paddings[0]; u8[1] += paddings[1]; u8[2] += paddings[2]; } } scale_x *= scale; scale_y *= scale; } /* the following code from OpenCV's haar feature implementation */ if(params.min_neighbors == 0) { for (i = 0; ENDORSE(i < seq->rnum); i++) { ccv_comp_t* comp = (ccv_comp_t*)ENDORSE(ccv_array_get(seq, i)); ccv_array_push(result_seq, comp); } } else { idx_seq = 0; ccv_array_clear(seq2); // group retrieved rectangles in order to filter out noise int ncomp = ccv_array_group(seq, &idx_seq, _ccv_is_equal_same_class, 0); ccv_comp_t* comps = (ccv_comp_t*)ccmalloc((ncomp + 1) * sizeof(ccv_comp_t)); memset(comps, 0, (ncomp + 1) * sizeof(ccv_comp_t)); // count number of neighbors for(i = 0; ENDORSE(i < seq->rnum); i++) { ccv_comp_t r1 = *(ccv_comp_t*)ENDORSE(ccv_array_get(seq, i)); int idx = *(int*)ENDORSE(ccv_array_get(idx_seq, i)); if (ENDORSE(comps[idx].neighbors) == 0) comps[idx].classification.confidence = r1.classification.confidence; ++comps[idx].neighbors; comps[idx].rect.x += r1.rect.x; comps[idx].rect.y += r1.rect.y; comps[idx].rect.width += r1.rect.width; comps[idx].rect.height += r1.rect.height; comps[idx].classification.id = r1.classification.id; comps[idx].classification.confidence = ccv_max(comps[idx].classification.confidence, r1.classification.confidence); } // calculate average bounding box for(i = 0; ENDORSE(i < ncomp); i++) { int n = ENDORSE(comps[i].neighbors); if(n >= params.min_neighbors) { ccv_comp_t comp; comp.rect.x = (comps[i].rect.x * 2 + n) / (2 * n); comp.rect.y = (comps[i].rect.y * 2 + n) / (2 * n); comp.rect.width = (comps[i].rect.width * 2 + n) / (2 * n); comp.rect.height = (comps[i].rect.height * 2 + n) / (2 * n); comp.neighbors = comps[i].neighbors; comp.classification.id = comps[i].classification.id; comp.classification.confidence = comps[i].classification.confidence; ccv_array_push(seq2, &comp); } } // filter out small face rectangles inside large face rectangles for(i = 0; ENDORSE(i < seq2->rnum); i++) { ccv_comp_t r1 = *(ccv_comp_t*)ENDORSE(ccv_array_get(seq2, i)); APPROX int flag = 1; for(j = 0; ENDORSE(j < seq2->rnum); j++) { ccv_comp_t r2 = *(ccv_comp_t*)ENDORSE(ccv_array_get(seq2, j)); APPROX int distance = (int)(r2.rect.width * 0.25 + 0.5); if(ENDORSE(i != j && r1.classification.id == r2.classification.id && r1.rect.x >= r2.rect.x - distance && r1.rect.y >= r2.rect.y - distance && r1.rect.x + r1.rect.width <= r2.rect.x + r2.rect.width + distance && r1.rect.y + r1.rect.height <= r2.rect.y + r2.rect.height + distance && (r2.neighbors > ccv_max(3, r1.neighbors) || r1.neighbors < 3))) { flag = 0; break; } } if(ENDORSE(flag)) ccv_array_push(result_seq, &r1); } ccv_array_free(idx_seq); ccfree(comps); } } ccv_array_free(seq); ccv_array_free(seq2); ccv_array_t* result_seq2; /* the following code from OpenCV's haar feature implementation */ if (params.flags & CCV_BBF_NO_NESTED) { result_seq2 = ccv_array_new(sizeof(ccv_comp_t), 64, 0); idx_seq = 0; // group retrieved rectangles in order to filter out noise int ncomp = ccv_array_group(result_seq, &idx_seq, _ccv_is_equal, 0); ccv_comp_t* comps = (ccv_comp_t*)ccmalloc((ncomp + 1) * sizeof(ccv_comp_t)); memset(comps, 0, (ncomp + 1) * sizeof(ccv_comp_t)); // count number of neighbors for(i = 0; ENDORSE(i < result_seq->rnum); i++) { ccv_comp_t r1 = *(ccv_comp_t*)ENDORSE(ccv_array_get(result_seq, i)); int idx = *(int*)ENDORSE(ccv_array_get(idx_seq, i)); if (ENDORSE(comps[idx].neighbors == 0 || comps[idx].classification.confidence < r1.classification.confidence)) { comps[idx].classification.confidence = r1.classification.confidence; comps[idx].neighbors = 1; comps[idx].rect = r1.rect; comps[idx].classification.id = r1.classification.id; } } // calculate average bounding box for(i = 0; ENDORSE(i < ncomp); i++) if(ENDORSE(comps[i].neighbors)) ccv_array_push(result_seq2, &comps[i]); ccv_array_free(result_seq); ccfree(comps); } else { result_seq2 = result_seq; } for (i = 1; ENDORSE(i < scale_upto + next * 2); i++) ccv_matrix_free(pyr[i * 4]); if (params.accurate) for (i = next * 2; ENDORSE(i < scale_upto + next * 2); i++) { ccv_matrix_free(pyr[i * 4 + 1]); ccv_matrix_free(pyr[i * 4 + 2]); ccv_matrix_free(pyr[i * 4 + 3]); } if (ENDORSE(params.size.height != _cascade[0]->size.height || params.size.width != _cascade[0]->size.width)) ccv_matrix_free(pyr[0]); return result_seq2; }
int main(int argc, char** argv) { static struct option swt_options[] = { /* help */ {"help", 0, 0, 0}, /* optional parameters */ {"size", 1, 0, 0}, {"low-thresh", 1, 0, 0}, {"high-thresh", 1, 0, 0}, {"max-height", 1, 0, 0}, {"min-height", 1, 0, 0}, {"min-area", 1, 0, 0}, {"aspect-ratio", 1, 0, 0}, {"std-ratio", 1, 0, 0}, {"thickness-ratio", 1, 0, 0}, {"height-ratio", 1, 0, 0}, {"intensity-thresh", 1, 0, 0}, {"letter-occlude-thresh", 1, 0, 0}, {"distance-ratio", 1, 0, 0}, {"intersect-ratio", 1, 0, 0}, {"letter-thresh", 1, 0, 0}, {"elongate-ratio", 1, 0, 0}, {"breakdown-ratio", 1, 0, 0}, {"breakdown", 1, 0, 0}, {"iterations", 1, 0, 0}, {"base-dir", 1, 0, 0}, {0, 0, 0, 0} }; if (argc <= 1) exit_with_help(); ccv_swt_param_t params = { .interval = 1, .same_word_thresh = { 0.2, 0.8 }, .min_neighbors = 1, .scale_invariant = 0, .size = 3, .low_thresh = 78, .high_thresh = 214, .max_height = 300, .min_height = 10, .min_area = 75, .letter_occlude_thresh = 2, .aspect_ratio = 10, .std_ratio = 0.5, .thickness_ratio = 1.5, .height_ratio = 2.0, .intensity_thresh = 45, .distance_ratio = 3.0, .intersect_ratio = 2.0, .letter_thresh = 3, .elongate_ratio = 1.3, .breakdown = 1, .breakdown_ratio = 1.0, }; ccv_swt_range_t size_range = { .min_value = 1, .max_value = 3, .step = 2, .enable = 1, }; ccv_swt_range_t low_thresh_range = { .min_value = 50, .max_value = 150, .step = 1, .enable = 1, }; ccv_swt_range_t high_thresh_range = { .min_value = 200, .max_value = 350, .step = 1, .enable = 1, }; ccv_swt_range_t max_height_range = { .min_value = 500, .max_value = 500, .step = 1, .enable = 1, }; ccv_swt_range_t min_height_range = { .min_value = 5, .max_value = 30, .step = 1, .enable = 1, }; ccv_swt_range_t min_area_range = { .min_value = 10, .max_value = 100, .step = 1, .enable = 1, }; ccv_swt_range_t letter_occlude_thresh_range = { .min_value = 0, .max_value = 5, .step = 1, .enable = 1, }; ccv_swt_range_t aspect_ratio_range = { .min_value = 5, .max_value = 15, .step = 1, .enable = 1, }; ccv_swt_range_t std_ratio_range = { .min_value = 0.1, .max_value = 1.0, .step = 0.01, .enable = 1, }; ccv_swt_range_t thickness_ratio_range = { .min_value = 1.0, .max_value = 2.0, .step = 0.1, .enable = 1, }; ccv_swt_range_t height_ratio_range = { .min_value = 1.0, .max_value = 3.0, .step = 0.1, .enable = 1, }; ccv_swt_range_t intensity_thresh_range = { .min_value = 1, .max_value = 50, .step = 1, .enable = 1, }; ccv_swt_range_t distance_ratio_range = { .min_value = 1.0, .max_value = 5.0, .step = 0.1, .enable = 1, }; ccv_swt_range_t intersect_ratio_range = { .min_value = 0.0, .max_value = 5.0, .step = 0.1, .enable = 1, }; ccv_swt_range_t letter_thresh_range = { .min_value = 0, .max_value = 5, .step = 1, .enable = 1, }; ccv_swt_range_t elongate_ratio_range = { .min_value = 0.1, .max_value = 2.5, .step = 0.1, .enable = 1, }; ccv_swt_range_t breakdown_ratio_range = { .min_value = 0.5, .max_value = 1.5, .step = 0.01, .enable = 1, }; int i, j, k, iterations = 10; while (getopt_long_only(argc - 1, argv + 1, "", swt_options, &k) != -1) { switch (k) { case 0: exit_with_help(); case 1: decode_range(optarg, &size_range); break; case 2: decode_range(optarg, &low_thresh_range); break; case 3: decode_range(optarg, &high_thresh_range); break; case 4: decode_range(optarg, &max_height_range); break; case 5: decode_range(optarg, &min_height_range); break; case 6: decode_range(optarg, &min_area_range); break; case 7: decode_range(optarg, &aspect_ratio_range); break; case 8: decode_range(optarg, &std_ratio_range); break; case 9: decode_range(optarg, &thickness_ratio_range); break; case 10: decode_range(optarg, &height_ratio_range); break; case 11: decode_range(optarg, &intensity_thresh_range); break; case 12: decode_range(optarg, &letter_occlude_thresh_range); break; case 13: decode_range(optarg, &distance_ratio_range); break; case 14: decode_range(optarg, &intersect_ratio_range); break; case 15: decode_range(optarg, &letter_thresh_range); break; case 16: decode_range(optarg, &elongate_ratio_range); break; case 17: decode_range(optarg, &breakdown_ratio_range); break; case 18: params.breakdown = !!atoi(optarg); break; case 19: iterations = atoi(optarg); break; case 20: chdir(optarg); break; } } FILE* r = fopen(argv[1], "rt"); if (!r) exit_with_help(); ccv_enable_cache(1024 * 1024 * 1024); ccv_array_t* aof = ccv_array_new(sizeof(char*), 64, 0); ccv_array_t* aow = ccv_array_new(sizeof(ccv_array_t*), 64, 0); ccv_array_t* cw = 0; char* file = (char*)malloc(1024); size_t len = 1024; ssize_t read; while ((read = getline(&file, &len, r)) != -1) { while(read > 1 && isspace(file[read - 1])) read--; file[read] = 0; double x, y, width, height; int recognized = sscanf(file, "%lf %lf %lf %lf", &x, &y, &width, &height); if (recognized == 4) { ccv_rect_t rect = { .x = (int)(x + 0.5), .y = (int)(y + 0.5), .width = (int)(width + 0.5), .height = (int)(height + 0.5) }; ccv_array_push(cw, &rect); } else { char* name = (char*)malloc(ccv_min(1023, strlen(file)) + 1); strncpy(name, file, ccv_min(1023, strlen(file)) + 1); ccv_array_push(aof, &name); cw = ccv_array_new(sizeof(ccv_rect_t), 1, 0); ccv_array_push(aow, &cw); } } free(file); printf("loaded %d images for parameter search of:\n", aof->rnum); if (size_range.enable) printf(" - canny size from %d to %d, += %lg\n", (int)(size_range.min_value + 0.5), (int)(size_range.max_value + 0.5), size_range.step); if (std_ratio_range.enable) printf(" - std threshold ratio from %lg to %lg, += %lg\n", std_ratio_range.min_value, std_ratio_range.max_value, std_ratio_range.step); if (max_height_range.enable) printf(" - maximum height from %d to %d, += %lg\n", (int)(max_height_range.min_value + 0.5), (int)(max_height_range.max_value + 0.5), max_height_range.step); if (min_height_range.enable) printf(" - minimum height from %d to %d, += %lg\n", (int)(min_height_range.min_value + 0.5), (int)(min_height_range.max_value + 0.5), min_height_range.step); if (min_area_range.enable) printf(" - minimum area from %d to %d, += %lg\n", (int)(min_area_range.min_value + 0.5), (int)(min_area_range.max_value + 0.5), min_area_range.step); if (letter_occlude_thresh_range.enable) printf(" - letter occlude threshold from %d to %d, += %lg\n", (int)(letter_occlude_thresh_range.min_value + 0.5), (int)(letter_occlude_thresh_range.max_value + 0.5), letter_occlude_thresh_range.step); if (aspect_ratio_range.enable) printf(" - aspect ratio threshold from %lg to %lg, += %lg\n", aspect_ratio_range.min_value, aspect_ratio_range.max_value, aspect_ratio_range.step); if (thickness_ratio_range.enable) printf(" - thickness ratio threshold from %lg to %lg, += %lg\n", thickness_ratio_range.min_value, thickness_ratio_range.max_value, thickness_ratio_range.step); if (height_ratio_range.enable) printf(" - height ratio threshold from %lg to %lg, += %lg\n", height_ratio_range.min_value, height_ratio_range.max_value, height_ratio_range.step); if (intensity_thresh_range.enable) printf(" - intensity threshold from %d to %d, += %lg\n", (int)(intensity_thresh_range.min_value + 0.5), (int)(intensity_thresh_range.max_value + 0.5), intensity_thresh_range.step); if (distance_ratio_range.enable) printf(" - distance ratio threshold from %lg to %lg, += %lg\n", distance_ratio_range.min_value, distance_ratio_range.max_value, distance_ratio_range.step); if (intersect_ratio_range.enable) printf(" - intersect ratio threshold from %lg to %lg, += %lg\n", intersect_ratio_range.min_value, intersect_ratio_range.max_value, intersect_ratio_range.step); if (letter_thresh_range.enable) printf(" - minimum number of letters from %d to %d, += %lg\n", (int)(letter_thresh_range.min_value + 0.5), (int)(letter_thresh_range.max_value + 0.5), letter_thresh_range.step); if (elongate_ratio_range.enable) printf(" - elongate ratio threshold from %lg to %lg, += %lg\n", elongate_ratio_range.min_value, elongate_ratio_range.max_value, elongate_ratio_range.step); if (breakdown_ratio_range.enable) printf(" - breakdown ratio threshold from %lg to %lg, += %lg\n", breakdown_ratio_range.min_value, breakdown_ratio_range.max_value, breakdown_ratio_range.step); if (low_thresh_range.enable) printf(" - canny low threshold from %d to %d, += %lg\n", (int)(low_thresh_range.min_value + 0.5), (int)(low_thresh_range.max_value + 0.5), low_thresh_range.step); if (high_thresh_range.enable) printf(" - canny high threshold from %d to %d, += %lg\n", (int)(high_thresh_range.min_value + 0.5), (int)(high_thresh_range.max_value + 0.5), high_thresh_range.step); double best_f = 0, best_precision = 0, best_recall = 0; double a = 0.5; double v; ccv_swt_param_t best_params = params; #define optimize(parameter, type, rounding) \ if (parameter##_range.enable) \ { \ params = best_params; \ int total_iterations = 0; \ for (v = parameter##_range.min_value; v <= parameter##_range.max_value; v += parameter##_range.step) \ ++total_iterations; \ double* precision = (double*)ccmalloc(sizeof(double) * total_iterations); \ double* recall = (double*)ccmalloc(sizeof(double) * total_iterations); \ double* total_words = (double*)ccmalloc(sizeof(double) * total_iterations); \ memset(precision, 0, sizeof(double) * total_iterations); \ memset(recall, 0, sizeof(double) * total_iterations); \ memset(total_words, 0, sizeof(double) * total_iterations); \ double total_truth = 0; \ for (j = 0; j < aof->rnum; j++) \ { \ char* name = *(char**)ccv_array_get(aof, j); \ ccv_dense_matrix_t* image = 0; \ ccv_read(name, &image, CCV_IO_GRAY | CCV_IO_ANY_FILE); \ ccv_array_t* truth = *(ccv_array_t**)ccv_array_get(aow, j); \ total_truth += truth->rnum; \ for (v = parameter##_range.min_value, k = 0; v <= parameter##_range.max_value; v += parameter##_range.step, k++) \ { \ params.parameter = (type)(v + rounding); \ ccv_array_t* words = ccv_swt_detect_words(image, params); \ double one_precision = 0, one_recall = 0; \ _ccv_evaluate_wolf(words, truth, params, &one_precision, &one_recall); \ assert(one_precision <= words->rnum + 0.1); \ precision[k] += one_precision; \ recall[k] += one_recall; \ total_words[k] += words->rnum; \ ccv_array_free(words); \ FLUSH("perform SWT on %s (%d / %d) for " #parameter " = (%lg <- [%lg, %lg])", name, j + 1, aof->rnum, v, parameter##_range.min_value, parameter##_range.max_value); \ } \ ccv_matrix_free(image); \ } \ for (v = parameter##_range.min_value, j = 0; v <= parameter##_range.max_value; v += parameter##_range.step, j++) \ { \ params.parameter = (type)(v + rounding); \ double f, total_precision = precision[j], total_recall = recall[j]; \ total_precision /= total_words[j]; \ total_recall /= total_truth; \ f = 1.0 / (a / total_precision + (1.0 - a) / total_recall); \ if (f > best_f) \ { \ best_params = params; \ best_f = f; \ best_precision = total_precision; \ best_recall = total_recall; \ } \ FLUSH("current harmonic mean : %.2lf%%, precision : %.2lf%%, recall : %.2lf%% ; best harmonic mean : %.2lf%%, precision : %.2lf%%, recall : %.2lf%% ; at " #parameter " = %lg (%lg <- [%lg, %lg])", f * 100, total_precision * 100, total_recall * 100, best_f * 100, best_precision * 100, best_recall * 100, (double)best_params.parameter, v, parameter##_range.min_value, parameter##_range.max_value); \ } \ printf("\n"); \ ccfree(precision); \ ccfree(recall); \ ccfree(total_words); \ } for (i = 0; i < iterations; i++) { optimize(size, int, 0.5); optimize(std_ratio, double, 0); optimize(max_height, int, 0.5); optimize(min_height, int, 0.5); optimize(min_area, int, 0.5); optimize(letter_occlude_thresh, int, 0.5); optimize(aspect_ratio, double, 0); optimize(thickness_ratio, double, 0); optimize(height_ratio, double, 0); optimize(intensity_thresh, int, 0.5); optimize(distance_ratio, double, 0); optimize(intersect_ratio, double, 0); optimize(letter_thresh, int, 0.5); optimize(elongate_ratio, double, 0); optimize(breakdown_ratio, double, 0); optimize(low_thresh, int, 0.5); optimize(high_thresh, int, 0.5); printf("At iteration %d(of %d) : best parameters for swt is:\n" "\tsize = %d\n" "\tlow_thresh = %d\n" "\thigh_thresh = %d\n" "\tmax_height = %d\n" "\tmin_height = %d\n" "\tmin_area = %d\n" "\tletter_occlude_thresh = %d\n" "\taspect_ratio = %lf\n" "\tstd_ratio = %lf\n" "\tthickness_ratio = %lf\n" "\theight_ratio = %lf\n" "\tintensity_thresh = %d\n" "\tdistance_ratio = %lf\n" "\tintersect_ratio = %lf\n" "\tletter_thresh = %d\n" "\telongate_ratio = %lf\n" "\tbreakdown_ratio = %lf\n", i + 1, iterations, best_params.size, best_params.low_thresh, best_params.high_thresh, best_params.max_height, best_params.min_height, best_params.min_area, best_params.letter_occlude_thresh, best_params.aspect_ratio, best_params.std_ratio, best_params.thickness_ratio, best_params.height_ratio, best_params.intensity_thresh, best_params.distance_ratio, best_params.intersect_ratio, best_params.letter_thresh, best_params.elongate_ratio, best_params.breakdown_ratio); } #undef optimize for (i = 0; i < aof->rnum; i++) { char* name = *(char**)ccv_array_get(aof, i); free(name); ccv_array_t* cw = *(ccv_array_t**)ccv_array_get(aow, i); ccv_array_free(cw); } ccv_array_free(aof); ccv_array_free(aow); ccv_drain_cache(); return 0; }