UVAGaze: Unsupervised 1-to-2 Views Adaptation for Gaze Estimation

Authors: Ruicong Liu, Feng Lu

AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental The experimental results show that a single-view estimator, when adapted for dual views, can achieve much higher accuracy, especially in cross-dataset settings, with a substantial improvement of 47.0%.
Researcher Affiliation Academia Ruicong Liu, Feng Lu* State Key Laboratory of VR Technology and Systems, School of CSE, Beihang University, Beijing, China {liuruicong, lufeng}@buaa.edu.cn
Pseudocode Yes Algorithm 1: Unsupervised 1-to-2 views adaptation. Input: Dual-camera input D, pre-training dataset Dpre and G pre-trained on Dpre Output: G( |Θ)
Open Source Code Yes Project page: https://github.com/ Mickey LLG/UVAGaze.
Open Datasets Yes We pre-train using two datasets: ETH-XGaze (Zhang et al. 2020) and Gaze360 (Kellnhofer et al. 2019). We have named the fine-tuned dataset as ETH-MV (Multi-View), which is the foundation for all our experiments. The ETH-MV is available at the project page.
Dataset Splits No The paper mentions pre-training datasets and dual-camera input for adaptation, and describes splitting cameras into pairs, but does not specify explicit train/validation/test splits (e.g., percentages or counts) for the data used in the adaptation process.
Hardware Specification Yes We use Py Torch on an NVIDIA 3090 GPU.
Software Dependencies No We use Py Torch on an NVIDIA 3090 GPU. The paper mentions PyTorch but does not specify its version or any other software dependencies with version numbers.
Experiment Setup Yes Pre-training employs the Adam optimizer at a learning rate of 10 4, while UVA uses 10 5. empirically, we set λ1 = 50 and λ2 = 10.