Bootstrap AutoEncoders With Contrastive Paradigm for Self-supervised Gaze Estimation
Authors: Yaoming Wang, Jin Li, Wenrui Dai, Bowen Shi, Xiaopeng Zhang, Chenglin Li, Hongkai Xiong
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that the proposed approaches outperform state-of-the-art unsupervised gaze approaches on extensive datasets (including wild scenes) under both within-dataset and cross-dataset protocols. |
| Researcher Affiliation | Collaboration | 1School of Electronic Information and Electrical Engineering, Shanghai Jiao Tong University, Shanghai, China 2Huawei Inc, Shenzhen, China. |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide an explicit statement or link for open-source code availability for the described methodology. |
| Open Datasets | Yes | Datasets. We perform our self-supervised experiments on 4 gaze datasets: Columbia Gaze (Smith et al., 2013), MPIIFace Gaze (Zhang et al., 2017b), Gaze360 (Kellnhofer et al., 2019) and ETH-Xgaze (Zhang et al., 2020). |
| Dataset Splits | Yes | Following the convention, a 5-fold evaluation protocol is adopted for Columbia. MPIIFace Gaze (MPII)...evaluated with a leave-one-out evaluation protocol. |
| Hardware Specification | Yes | We implement the codes with the Pytorch (Paszke et al., 2019) framework and use 4 Nvidia-V100 GPUs for training. |
| Software Dependencies | No | We implement the codes with the Pytorch (Paszke et al., 2019) framework". While PyTorch is mentioned with a citation, a specific version number is not provided, nor are versions for other key software components. |
| Experiment Setup | Yes | An Adam W optimizer and a cosine decay learning rate schedule are used with the initial learning rate settled as 4 10 4 for Res Net and 1.5 10 4 for Vi T-tiny. A 0.05 weight-decay is also employed and we warm up the training process with 10 epochs and then train the model for 190 epochs (The total epochs are 200). |