Chengan Che, Chao Wang, Xinyue Chen, Sophia Tsoka, Luis C. Garcia-Peraza-Herrera
CVPR 2026
PL-Stitch learns procedure-aware video representations by formulating temporal ordering as a probabilistic listwise ranking problem. It combines global temporal ranking with a spatio-temporal jigsaw objective to capture both procedural progression and fine-grained temporal correspondence.
]]>
Chengan Che, Chao Wang, Tom Vercauteren, Sophia Tsoka, Luis C. Garcia-Peraza-Herrera
CVPR 2026
LEMON is a large-scale open-access surgical video dataset comprising 4,194 long-form videos, approximately 938 hours of footage, and 85 million frames across 35 surgical procedures.
To curate this dataset, we developed a highly automated pipeline for collecting, filtering, and processing unconstrained surgical videos. We further pretrained LemonFM on LEMON using an augmented self-distillation framework, producing transferable visual representations for surgical video understanding.
LemonFM achieves strong performance across multiple downstream tasks, including surgical phase recognition, tool and action recognition, and semantic segmentation.
]]>
Chengan Che, Chao Wang, Jiayuan Huang, Xinyue Chen, Luis C. Garcia-Peraza-Herrera
CVPR Workshops 2026, Oral Presentation
SurgLIME enables scalable surgical vision-language pre-training using LLM-generated textual descriptions. It combines parameter-efficient vision-language alignment with a confidence-aware mechanism that reduces the contribution of less reliable generated text during training.
* Equal contribution.
]]>
Chao Wang, Chengan Che, Xinyue Chen, Sophia Tsoka, Luis C. Garcia-Peraza-Herrera
CVPR 2026
Back to the Feature explains video classifiers by generating counterfactual examples in feature space. The method identifies changes that alter model predictions while preserving the temporal and semantic structure of the original video.
]]>