This repository provides the implementation of the VRDFormer: End-to-end Video Relation Detection with Transformers paper. The codebase builds upon DETR, Deformable DETR and TrackFormer.
Visual relation understanding plays an essential role for holistic video understanding. Most previous works adopt a multi-stage framework for video visual relation detection (VidVRD), which cannot capture long-term spatiotemporal contexts in different stages and also suffers from inefficiency. In this paper, we propose a transformerbased framework called VRDFormer to unify these decoupling stages. Our model exploits a query-based approach to autoregressively generate relation instances. We specifically design static queries and recurrent queries to enable efficient object pair tracking with spatio-temporal contexts. The model is jointly trained with object pair detection and relation classification. Extensive experiments on two benchmark datasets, ImageNet-VidVRD and VidOR, demonstrate the effectiveness of the proposed VRDFormer, which achieves the state-of-the-art performance on both relation detection and relation tagging tasks.
We refer to our docs/DATA.md for detailed installation instructions.
We refer to our docs/INSTALL.md for detailed installation instructions.
- Train VidVRD based on detr with 8GPUs and batchsize=32
sh script/stage1/train_mgpu.sh
sh script/stage2/train_mgpu.sh
- Train VidVRD based on deformable detr
sh script/stage1/train_deform_mgpu.sh
sh script/stage2/train_deform_mgpu.sh
Train VidOR based on detr with 8GPUs and batchsize=64
sh script/stage1/train_mgpu_vidor.sh
sh script/stage2/train_mgpu_vidor.sh
Train VidOR based on deformable detr
sh script/stage1/train_deform_mgpu_vidor.sh
sh script/stage2/train_deform_mgpu_vidor.sh