Gaze Label Alignment: Alleviating Domain Shift for Gaze Estimation
Authors: Guanzhong Zeng, Jingjing Wang, Zefu Xu, Pengwei Yin, Wenqi Ren, Di Xie, Jiang Zhu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our GLA method can effectively alleviate the label distribution shift, and SOTA gaze estimation methods can be further improved obviously. Experiments Data Preparation Due to the wide range of gaze angle, ETH-XGaze (Zhang et al. 2020) and Gaze360 (Kellnhofer et al. 2019) are commonly used as training datasets. |
| Researcher Affiliation | Industry | Hikvision Research Institute, Hangzhou, China EMAIL |
| Pseudocode | No | The paper describes the method in a step-by-step manner (Step1: Eliminate the data distribution shift., Step2: Train gaze regressor on selected domain., etc.) but does not use a structured pseudocode block or algorithm environment. |
| Open Source Code | No | The paper states "More implementation details can be found in the supplementary material." but does not explicitly mention releasing source code, nor does it provide a link to a repository. |
| Open Datasets | Yes | Due to the wide range of gaze angle, ETH-XGaze (Zhang et al. 2020) and Gaze360 (Kellnhofer et al. 2019) are commonly used as training datasets. Additionally, we add Gaze Capture (Krafka et al. 2016) to construct more source domains. To maintain consistency with previous works (Cheng and Bao 2022; Wang et al. 2022b; Cai et al. 2023), we select MPIIFace Gaze (Zhang et al. 2017) and Eye Diap (Funes Mora, Monay, and Odobez 2014) as our target datasets. |
| Dataset Splits | No | The paper mentions "To avoid the impact of data imbalance, we resample data to maintain the same amount of data for each domain." and "We randomly choose 100 samples from target domain and report average results of 20 repeated trials" for unsupervised domain adaptation, but does not provide general training/validation/test splits for all datasets or experiments. |
| Hardware Specification | No | The paper does not provide any specific hardware details such as GPU models, CPU types, or memory amounts used for the experiments. |
| Software Dependencies | No | The paper mentions using "ResNet-18 (He et al. 2016)", "SGD Optimizer", "a two layers MLP regressor", and "Cosineannealing LR scheduler (Loshchilov and Hutter 2016)", but does not provide specific version numbers for any software or libraries. |
| Experiment Setup | Yes | We employ ResNet-18 (He et al. 2016) as our gaze feature extractor F, a two layers MLP regressor W to predict gaze angle. For baseline, we employ SGD Optimizer with a learning rate of 5e-2. The batch size is 126, and all images are resized to 224x224. We train the gaze estimation model for 30 epochs utilizing a Cosineannealing LR scheduler (Loshchilov and Hutter 2016) with a 3 epoch warm-up. We perform a data augmentation family with a random color field and greyscale like (Wang et al. 2022b). |