dataset-movi_c

This is dataset MOVi-C in LMDB database format, which can be used off-the-shelf in this repo.

archive-videosaur

Here are model checkpoints for baseline VideoSAUR.
Using DINO2 S/14 for encoding.
Models are trained on datasets MOVi-C/D and YTVIS (high quality), with random seeds 42, 43 and 44.
Input resolution is 256x256 (224x224).

archive-slotcontrast

Here are model checkpoints for baseline SlotContrast.
Using DINO2 S/14 for encoding.
Models are trained on datasets MOVi-C/D and YTVIS (high quality), with random seeds 42, 43 and 44.
Input resolution is 256x256 (224x224).

archive-recogn

Here are model checkpoints for object recognition models powered by RandSF.Q-tsim and SlotContrast.
Using DINO2 S/14 for encoding.
Models are trained on dataset YTVIS (high quality), with random seeds 42, 43 and 44.
Input resolution is 256x256 (224x224).
Slot matching threshold is 1e-1@IoU.

archive-randsfq-tsim

Here are model checkpoints for our RandSF.Q, built upon SlotContrast, but using time similarity loss.
Using DINO2 S/14 for encoding.
Models are trained on datasets MOVi-C/D and YTVIS (high quality), with random seeds 42, 43 and 44.
Input resolution is 256x256 (224x224).

archive-randsfq

Here are model checkpoints for our RandSF.Q, built upon SlotContrast.
Using DINO2 S/14 for encoding.
Models are trained on datasets MOVi-C/D and YTVIS (high quality), with random seeds 42, 43 and 44.
Input resolution is 256x256 (224x224).

dataset-ytvis

This is dataset YTVIS in LMDB database format, which can be used off-the-shelf in this repo.