PushUpBench: Video Repetition Counting Benchmark

Project Page | Paper | GitHub

PushUpBench is a benchmark for evaluating vision-language models (VLMs) on their ability to count exercise repetitions in videos. It was introduced in the paper "PushupBench: Your VLM is not good at counting pushups". The dataset consists of 446 long-form clips (averaging 36.7s) designed to test temporal reasoning and repetition counting beyond simple pattern recognition.

Dataset Structure

Each sample contains:

• name: Action description (e.g., "push ups", "leg lift", "knee to chest")
• video_path: Filename of the video
• count: List of acceptable count values (some exercises have ambiguous boundaries)
• fuzzy_action: Whether the action has ambiguous start/end boundaries
• complex_action: Whether the action is compound/complex

Usage with lmms-eval

PushUpBench is incorporated in the lmms-eval toolkit.

# Set the video directory
export PUSHUPBENCH_VIDEO_DIR=/path/to/videos

# Run evaluation
python -m lmms_eval \
    --model <model> \
    --tasks pushupbench \
    --batch_size 1 \
    --output_path results/

Metrics

• Exact Match: Prediction matches any value in the ground truth count list.
• MAE: Mean Absolute Error between prediction and primary ground truth.
• OBO: Off-By-One accuracy (prediction within 1 of any ground truth).