metadata
task_categories:
- video-text-to-text
VisCoP Dataset
This repository contains the training and evaluation data for VisCoP: Visual Probing for Video Domain Adaptation of Vision Language Models.
- Paper: VisCoP: Visual Probing for Video Domain Adaptation of Vision Language Models
- Repository: GitHub - dominickrei/VisCoP
Dataset Description
VisCoP is a parameter-efficient adaptation framework designed to adapt Vision Language Models (VLMs) to new domains (e.g., cross-view, cross-modal, and cross-task settings) using a compact set of learnable visual probes. This dataset provides the instruction pairs and videos used for training and evaluating VisCoP across these scenarios.
Training Data
The training dataset includes:
- Egocentric Viewpoint: Instructions and videos.
- Depth Modality: Instructions and videos.
Evaluation Data
We evaluate VisCoP across multiple target domains using the following benchmarks:
- Egocentric Viewpoint: Ego-in-Exo PerceptionMCQ, EgoSchema, NeXTQA, VideoMME, ADL-X.
- Depth Modality: Exo Depth videos (contained in
depth_videos.zip). - Robot Control: VIMA-Bench.
For detailed setup and evaluation protocols, please refer to the GitHub Repository.
Citation
@inproceedings{reilly2026viscop,
title = {VisCoP: Visual Probing for Video Domain Adaptation of Vision Language Models},
author = {Dominick Reilly and Manish Kumar Govind and Le Xue and Srijan Das},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
year = {2026}
}