>From 499de7be409ba7668defac9214a8343344bd21ee Mon Sep 17 00:00:00 2001 From: MaximilianKaindl Date: Wed, 12 Mar 2025 10:10:58 +0100 Subject: [PATCH v2 FFmpeg 21/20] libavfilter: classify fix category post_processing with very low temperature To: ffmpeg-devel@ffmpeg.org Signed-off-by: MaximilianKaindl --- libavfilter/avf_dnn_classify.c | 24 ++++++++++++++---------- 1 file changed, 14 insertions(+), 10 deletions(-) diff --git a/libavfilter/avf_dnn_classify.c b/libavfilter/avf_dnn_classify.c index 7b469e3af0..ade55858ad 100644 --- a/libavfilter/avf_dnn_classify.c +++ b/libavfilter/avf_dnn_classify.c @@ -34,6 +34,7 @@ #include "libavutil/opt.h" #include "libavutil/time.h" #include "video.h" +#include /* Labels that are being used to classify the image @@ -579,7 +580,7 @@ static int get_category_total_label_count(CategoryClassifcationContext *cat_ctx, static CategoryContext *get_best_category(CategoriesContext *categories_ctx, float *probabilities) { CategoryContext *best_category = NULL; - float best_probability = -1.0f; + float best_probability = FLT_MIN; int prob_offset = 0; // Calculate total probability for each category @@ -592,7 +593,7 @@ static CategoryContext *get_best_category(CategoriesContext *categories_ctx, flo category->total_probability += probabilities[prob_offset + label_idx]; } - if (category->total_probability > best_probability) { + if (category->total_probability > best_probability && category->total_probability > 0.0f) { best_probability = category->total_probability; best_category = category; } @@ -761,7 +762,8 @@ static int post_proc_clxp_categories(AVFrame *frame, DNNData *output, uint32_t b av_freep(&ctx_labels); return AVERROR(ENOMEM); } - + + int category_count = 0; // Process each context for (int ctx_idx = 0; ctx_idx < cat_class_ctx->num_contexts; ctx_idx++) { CategoriesContext *categories_ctx = cat_class_ctx->category_units[ctx_idx]; @@ -773,20 +775,22 @@ static int post_proc_clxp_categories(AVFrame *frame, DNNData *output, uint32_t b // Find best category in context best_category = get_best_category(categories_ctx, probabilities + prob_offset); if (!best_category || !best_category->name) { - av_log(filter_ctx, AV_LOG_ERROR, "Invalid best category at context %d\n", ctx_idx); + // No category classification found continue; } // Copy category name instead of assigning pointer - av_strlcpy(ctx_labels[ctx_idx], best_category->name, AV_DETECTION_BBOX_LABEL_NAME_MAX_SIZE); - ctx_probabilities[ctx_idx] = best_category->total_probability; + av_strlcpy(ctx_labels[category_count], best_category->name, AV_DETECTION_BBOX_LABEL_NAME_MAX_SIZE); + ctx_probabilities[category_count] = best_category->total_probability; prob_offset += categories_ctx->label_count; + category_count++; + } + if(category_count > 0){ + // Fill bbox with best categories + ret = fill_detection_bbox_with_best_labels(ctx_labels, ctx_probabilities, cat_class_ctx->num_contexts, bbox, + AV_NUM_DETECTION_BBOX_CLASSIFY, ctx->confidence); } - - // Fill bbox with best labels - ret = fill_detection_bbox_with_best_labels(ctx_labels, ctx_probabilities, cat_class_ctx->num_contexts, bbox, - AV_NUM_DETECTION_BBOX_CLASSIFY, ctx->confidence); // Clean up for (int i = 0; i < cat_class_ctx->num_contexts; i++) { -- 2.34.1