DavidRobinson05/strokevision-bench

--- license: cc-by-nc-4.0 task_categories: - video-classification language: - en tags: - code pretty_name: StrokeVision Bench size_categories: - 1K<n<10K --- # StrokeVision-Bench: A Multimodal Video and 2D Pose Benchmark for Tracking Stroke Recovery StrokeVision-Bench is an action recognition dataset of short segments of stroke patients performing the Box-Block Test. StrokeVision-Bench contains 1,000 annotated videos (1 s @ 30 FPS) categorized into four clinically meaningful action classes (Non-task movement, Grasping, Transport with block, Transport without block), with each sample represented in two modalities: raw video segments and 2D skeletal keypoints. We benchmark several state-of-the-art video and skeleton-based action classification methods to establish performance baselines for this domain and facilitate future research in automated stroke rehabilitation assessment. ## Dataset Summary - Samples: 1,036 short videos (1 s @ 30 FPS) - Modalities: RGB frames, 2D skeleton keypoints - Action classes: Non-task movement, Grasping, Transport with block, Transport without block - Keypoints: COCO 17-keypoint format - Train-Test Split: 827 train segments, 209 test segments **Paper**: https://arxiv.org/abs/2509.07994 ## Dataset Structure - videos - grasping/ - non_task/ - transport_with_block/ - transport_without_block/ - keypoints/ - grasping/ - non_task/ - transport_with_block/ - transport_without_block/ - annotations/ - train.csv - val.csv The `videos` folder contains the raw video segments separated by class. The `keypoints` folder contains the 2D skeletal keypoints as npy files with shape (30, 17, 2) separated by class. Each instance in the annotations contains the following features: - subject_id: The subject of the instance (P01-P04) - file_name: File name of the instance within the **videos** and **keypoints** directories, formatted as "{subject_id}_segment{segment_id}" - label: Action class (Non-task movement, Grasping, Transport with block, Transport without block) - hand: Which hand is being used (left, right) Example entry: P01,P01_segment0201,transport_with_block,left You can load the annotations directly with pandas and access files via the dataset's `videos` and `keypoints` folders. ## License This dataset is released under the CC BY-NC 4.0 license. ## Citation ```bibtex @inproceedings{strokevisionbench, title = {StrokeVision-Bench: A Multimodal Video and 2D Pose Benchmark for Tracking Stroke Recovery}, author = {David Robinson and Animesh Gupta and Rizwan Qureshi and Qiushi Fu and Mubarak Shah}, booktitle = {Proceedings of the IEEE International Workshop on Machine Learning for Signal Processing (MLSP)}, year = {2025} } ```StrokeVision-Bench is an action recognition dataset of short segments of stroke patients performing the Box-Block Test, containing 1,000 annotated videos (1 s @ 30 FPS) categorized into four clinically meaningful action classes: Non-task movement, Grasping, Transport with block, Transport without block. Each sample is represented in two modalities: raw video segments and 2D skeletal keypoints. The dataset provides benchmarking for several state-of-the-art video and skeleton-based action classification methods to establish performance baselines for this domain and to facilitate future research in automated stroke rehabilitation assessment.