Single-Stage Visual Query Localization in Egocentric Videos

Hanwen Jiang, Santhosh Ramakrishnan, Kristen Grauman

conda create --name vqloc python=3.8
conda activate vqloc

# Install pytorch or use your own torch version
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.6 -c pytorch -c conda-forge

pip install -r requirements.txt 

We provide the model weights trained on here.

• Please follow vq2d baseline step 1/2/4/5 to process the dataset into video clips.

• Use ./train.sh and change your training config accordingly.
• The default training configurations require about 200GB at most, e.g. 8 A40 GPUs with 40GB VRAM, each.

• 1. Use ./inference_predict.sh to inference on the target video clips. Change the path of your model checkpoint.
• 2. Use python inference_results.py --cfg ./config/val.yaml to format the results. Use --eval and --cfg ./config/eval.yaml for evaluation (submit to leaderboard).
• 3. Use python evaluate.py to get the numbers. Please change --pred-file and --gt-file accordingly.

• The hard negative mining is not very steady. We set use_hnm=False by default.

@article{jiang2023vqloc,
   title={Single-Stage Visual Query Localization in Egocentric Videos},
   author={Jiang, Hanwen and Ramakrishnan, Santhosh and Grauman, Kristen},
   journal={ArXiv},
   year={2023},
   volume={2306.09324}
}