Self-supervised spatial features extracted from the MIPO dataset videos

This record contains precomputed self-supervised spatial feature representations extracted from videos of distal radius fracture surgery used in the study: C. Graëff, N. Padoy, P. Liverneaux, and T. Lampert, “Introducing surgical workflow recognition in orthopaedic surgery with timestamp supervision,” Computers in Biology and Medicine, 2025. These features correspond to the inputs used in the main experiments of the paper.

The self-supervised features were obtained with the DINO method using the code provided by Ramesh et al. at https://github.com/CAMMA-public/SelfSupSurg.

One file is provided for each video of the dataset used (MIPO dataset). Each file contains a matrix of shape (N, 2048), where N is the number of video frames sampled at 5 frames per second, and each row corresponds to a 2048-dimensional spatial feature vector.

For instructions on how these features can be used, please refer to the associated GitHub repository: https://github.com/camGcam/UCATD.