GCDR-Gaze

Overview

This is the official implementation of our paper: 'Diffusion-Refined VQA Annotations for Semi-Supervised Gaze Following' published in ECCV 2024 [paper]

We propose the first semi-supervised gaze following method, by leveraging VQA priors with Grad-CAM to generate pseudo annotations, and using a diffusion annotation prior to refine the initial noisy Grad-CAM heatmaps. The overall pipeline is shown in the above figure.

Environment

We used python=3.10, torch=2.1, and cuda-runtime=11.8 in our experiments.

Dataset

We used GazeFollow and VideoAttentionTarget datasets in our experiments.

We extended the annotations by detecting body boxes for the datasets. We also provide the randomly splitted annotations for semi-supervised training in GazeFollow with different amount of annotations. (GazeFollow. VideoAttentionTarget)

Grad-CAM Heatmap Generation

We use OFA as our VQA model. We modified the original code base to compute the Grad-CAM heatmap at the step when the model generates the key noun in its answer to gaze following question. The Grad-CAM heatmap is computed on the cross-attention layer in the transformer decoder of the VQA model.

The pretrained weights for the VQA model can be downloaded here.

Run Grad-CAM heatmap generation:

cd GradCAM_Compute
python gradcam_gen_largescale.py

We also provide the pre-computed Grad-CAM heatmaps here.

Semi-supervised learning with Diffusion Model Refinement

We train a diffusion model (conditioned with features from scene and head images) to refine the initial noisy Grad-CAM heatmaps into high-quality pseudo annotations for the gaze target.

Training teacher models on supervised data

Train diffusion model on labeled data (We trained 150 epochs with 5% and 10% labels, and 100 epochs for 20% labels), the feature extraction module of diffusion model is initialized same as VAT (link):

python train_diff_gazefollow.py --device {} --inference_steps 10 --time_steps 500 --onlyin --lr 5e-5 --batch_size 48 --supervise_ratio {0.05/0.1/0.2} --scale_input --scale 2.0 --epochs {} --eval_every 2

The diffusion model can also be enhanced by training with unlabeled data using the Mean Teacher method (We trained for 40 epochs for 5% labels, 60 epochs for 10%, and 80 epochs for 20%):

python train_diff_gazefollow_meanteacher.py  --device {} --consistency_type kl --no_inout --supervise_ratio {} --ema-decay 0.99 --batch_size 40 --onlyin --labeled-batch-size 10 --lr 5e-5 --inference_steps 10 --time_steps 500 --batch_size 40 --scale_input --scale 2.0 --consistency 4000 --epochs {} --eval_every 2

We also provide the code for training baseline teacher models (VAT, VAT-GC):

python train_gazefollow_base.py --device {} --batch_size 48 --epochs 40 --no_inout --lr 2.5e-4 --supervise_ratio {} --model {} --eval_every 2

Semi-supervised training with the trained diffusion model

 python train_gazefollow_semisup.py --device {gpuid} --teacher_model diffusion --teacher_ckpt {weights for trained teacher model} --time_steps 500 --supervise_ratio {0.05/0.1/0.2} --no_inout --onlyin --labeled_batch_size 20 --batch_size 80 --scale_input --scale 2.0  --model baseline --denoise_steps {} --inference_steps 2 

We set the "denoise_steps" argument as 200 when training with 5% labels, and 250 for other ratios.

You can test the performance for other baseline teacher models by changing the "teacher_model" argument.

We have also provided exemplary teacher model weights trained with 10% labels here.

Citation

If you find our method/code useful, please cite as follow:

@article{miao2024diffusion,
  title={Diffusion-Refined VQA Annotations for Semi-Supervised Gaze Following},
  author={Miao, Qiaomu and Graikos, Alexandros and Zhang, Jingwei and Mondal, Sounak and Hoai, Minh and Samaras, Dimitris},
  journal={arXiv preprint arXiv:2406.02774},
  year={2024}
}

Contact

If you have any questions, please email Qiaomu Miao at [email protected].