VD-IT model
The is our pre-trained checkpoint for our paper Exploring Pre-trained Text-to-Video Diffusion Models for Referring Video Object Segmentation.
We use a video diffusion model (ModelScopeT2V) as our base model, applying prompt tuning to adapt it as a visual backbone for downstream video understanding tasks.
Model traning
We first pre-train our model on Ref-COCO and then fine-tune it on Ref-YouTube-VOS. The training of the models utilizes two NVIDIA A100 GPUs, processing 5 frames per clip over the course of 9 epochs. The initial learning rate is set to 5e-5 and reduced by a factor of 10 at the 6th and 8th epochs.