DG-Net
Accepted at ACM Multimedia 2026
This repository provides the released checkpoint for the depth-guided video instance counting framework introduced in Depth-Guided Video Object Counting in Crowded Scenes. The model is designed for video instance counting in crowded and occluded RGB-D scenes.
Paper: Depth-Guided Video Object Counting in Crowded Scenes (ACM Multimedia 2026)
Model Description
DG-Net jointly uses RGB appearance and depth cues to improve object detection, occlusion estimation, cross-frame association, and video-level de-duplication. It supports text prompts and visual exemplars for open-vocabulary video counting.
Files
DG-Net/
|- DG-Net.pth
|- checksums.txt
|- LICENSE
`- README.md
DG-Net.pth: Released evaluation checkpoint.checksums.txt: SHA256 checksum for verifying file integrity.
Requirements
The model was tested on Linux with Python 3.10.18, CUDA Toolkit 12.8, GCC 11.3, and an NVIDIA GeForce RTX 5090 GPU. Please use the official code repository to install the required environment and build the Multi-Scale Deformable Attention CUDA extension.
Usage
Download DG-Net.pth and place it in the local checkpoints/ directory. Then run evaluation from the official code repository:
python scripts/evaluate.py \
--weights_path checkpoints/DG-Net.pth \
--config configs/config_test.py \
--split test \
--data_root data/Countdata \
--output_json_path results/test_with_visual_exemplars.json \
--use_exemplar_infer \
--exemplars_file data/exemplars_test.json
Training Data
The checkpoint is trained on RGBD-VideoCount and initialized from CountGD. Users must separately obtain the dataset, BERT files, and CountGD initialization checkpoint, and comply with their respective licenses and terms.
Limitations
This model is designed for crowded RGB-D video counting. Its performance may degrade when depth maps are inaccurate or unavailable, object categories are substantially different from the training data, or scenes contain severe motion blur and ambiguous object boundaries.
License
The released checkpoint is provided under the MIT License. The training dataset and third-party dependencies may be subject to separate license terms.
Citation
If you use this code, please cite our paper:
@inproceedings{xu2026depth,
title = {Depth-Guided Video Object Counting in Crowded Scenes},
author = {Xu, Yuanjing and Liu, Xinyan and Chen, Weidong and Zou, Zixuan and Zhang, Linhao and Meng, Zhuangzhe and Chan, Antoni B. and Zhang, Weigang},
booktitle = {Proceedings of the 34th ACM International Conference on Multimedia},
year = {2026},
doi = {10.1145/3767308.3835482}
}