ST-Evidence-Bench

ST-Evidence Benchmark Dataset ST-Evidence is a comprehensive benchmark for evaluating Spatial-Temporal Evidence generation in video understanding. It contains two tasks: **Generation (Gen)** and **Multiple Choice Question (MCQ)**. This was released for research purposes only, in support of the academic paper Evidence-Backed Video Question Answering. Dataset Overview **Total Videos**: ~1,300 videos at 6fps **Annotations**: Question-Answer pairs with temporal segments and spatial masks **Tasks**: Generation and MCQ **Domains**: Diverse video content Files Structure Generation Task (st_evidence_gen) Data Format st_evidence_gen.csv** contains the following columns: `entry_id`: Unique identifier for each question `video_id`: Video identifier `video_path`: Relative path to video file `question`: Question text `candidates`: List of answer options (for reference) `answer`: Ground truth answer `segment`: Temporal evidence segments [[start1, end1], [start2, end2], ...] Usage Evaluation Metrics **QA Accuracy**: Percentage of correct answers **Temporal IoU**: Intersection over Union for temporal segments mIoU, TIoU@0.3, TIoU@0.5 **Temporal IoP**: Intersection over Prediction mIoP, TIoP@0.3, TIoP@0.5 **Spatial Quality** (if masks generated): J score (Jaccard/IoU) F score (contour-based) J&F score (average) Distribution clevrer: 366 (13.53%) pt: 782 (28.90%) qa_ego4d_trimmed: 270 (9.98%) star: 327 (12.08%) vicas: 961 (35.51%) MCQ Task (st_evidence_mcq) Data Format st_evidence_mcq.csv** contains: `entry_id`: Unique identifier `video_id`: Video identifier `video_path`: Path to video `question`: Question text `candidates`: Answer options `answer`: Correct answer `segment`: Temporal evidence `mask_options`: Reference to mask options `temp_options`: Reference to temporal options mask_options.json**: Contains spatial mask options for each question temp_options.json**: Contains temporal segment options for each question options.tar.gz**: Pre-rendered mask visualizations for options (1,2