Here are model checkpoints and training logs for our RandSF.Q + ICC.
Using DINO2 S/14 for encoding.
Models are trained on datasets MOVi-C/E, YTVIS-HQ, with random seeds 42, 43 and 44, for both object discovery and recognition.
Input resolution is 256x256 (224x224).