abhrac/xmodal-vit

Official implementation of "Cross-Modal Fusion Distillation for Fine-Grained Sketch-Based Image Retrieval", BMVC 2022.

Cross-Modal Fusion Distillation for Fine-Grained Sketch-Based Image Retrieval

Official implementation of "Cross-Modal Fusion Distillation for Fine-Grained Sketch-Based Image Retrieval", BMVC 2022.
Additional Links: arXiv | Video & Poster

Our framework retains semantically relevant modality-specific features by learning a fused representation space, while bypassing the expensive cross-attention computation at test-time via cross-modal knowledge distillation.

Environment Setup

This project is implemented using PyTorch. A conda environment with all related dependencies can be created as follows:

1. Clone the project repository:

git clone https://github.com/abhrac/xmodal-vit.git
cd xmodal-vit

2. Create and activate conda environment:

conda env create -f environment.yml
conda activate xmodal-vit

Experimentation

To run the whole train-test pipeline end-to-end, run:

./run_expt.sh

Training

To train individual components from scratch, run the following:

python src/train_teacher.py --dataset=DatasetName
python src/train_photo_student.py --dataset=DatasetName
python src/train_sketch_student.py --dataset=DatasetName

where DatasetName is one of ShoeV2, ChairV2 or Sketchy.

Evaluation

Pre-trained models are available here. To evaluate a trained model, run:

python src/test.py --dataset=DatasetName

Results

	Shoe-V2	Shoe-V2	Chair-V2	Chair-V2
	Acc@1	Acc@10	Acc@1	Acc@10
Yang et al., ICCV '21	32.33	79.63	52.89	94.88
Sain et al., CVPR '21	36.47	81.83	62.86	91.14
Bhunia et al., CVPR '21	39.10	87.50	62.20	90.80
Chowdhury et al., CVPR '22	39.90	82.90	-	-
Bhunia et al., CVPR '22	43.70	-	64.80	-
Ours (XModalViT)	45.05	90.23	63.48	95.02

	Sketchy	Sketchy
	Acc@1	Acc@10
Human (Sangkloy et al., SIGGRAPH'16)	54.27	-
Pang et al., BMVC'17	50.14	-
Wang et al., PR'20 (S+I)	40.16	92.00
Wang et al., PR'20 (S+I+D)	46.20	96.49
Ours (XModalViT)	56.15	96.86

Citation

@inproceedings{Chaudhuri2022XModalViT,
 author = {Abhra Chaudhuri and Massimiliano Mancini and Yanbei Chen and Zeynep Akata and Anjan Dutta},
 booktitle = {Proceedings of the British Machine Vision Conference (BMVC)},
 title = {Cross-Modal Fusion Distillation for Fine-Grained Sketch-Based Image Retrieval},
 year = {2022}
}

xmodal-vit's Issues

Cannot reproduce the results

Hello！ I am very interested in your work. Here are the problems I cannot solve:
(1) I used your pretrianed model, but the results are not the same as in your paper, and also is the test.py wrong? the test.py is not test for all datasets.
(2) I used the RTX 4090 to train the model, and the parameters were set as yours, but the results are also worse, I felt a little confused.
I hope you'll be able to solve my current problems.
Thanks!