Projects & Datasets

Projects & code

• Training GRU: GRU-based cricket stroke localization.
• Finetuning C3D (Private): fine-tuning C3D for stroke localization.
• KTH Optical Flow: action recognition from optical flow visualizations.

LNMIIT Cricket Strokes Dataset

Benchmark datasets for temporal Cricket stroke localization in Highlight and Generic videos. Contact: arpan@lnmiit.ac.in

Highlights Dataset

• Description: The dataset has 26 untrimmed highlight videos of ICC World Cup T20 2016 matches, each of 3-5 min. duration with frame size of 360x640 at 25 FPS. The first 16 videos (when sorted by file name) are training videos, next 5 are validation videos and remaining 5 are testing videos. The annotations are provided separately as json files, with one json file corresponding to each video. The annotations are list of tuples (starting frame no., ending frame no.) of Cricket strokes.
• Links: Dataset | Labels | C3D Feats

Generic Dataset

• Description: The dataset has separate three partitions, training, validation and testing, with videos taken from ODI, Test and T20 matches. The videos were downsampled to 25 FPS and having 360x640 resolution. The total number of frames are approx. 1.93M.
• Training set: Dataset | Labels | C3D Feats
• Validation set: Dataset | Labels | C3D Feats
• Test set: Dataset | Labels | C3D Feats

Papers to cite

• Arpan Gupta, and Sakthi Balan Muthiah, "Viewpoint constrained and unconstrained Cricket stroke localization from untrimmed videos," Image and Vision Computing (Volume 100), August 2020. [DOI]
• Arpan Gupta, and Sakthi Balan M., "Temporal Cricket Stroke Localization from Untrimmed Highlight Videos," ICVGIP 2018. [ACM Link]