Boosting Image-based Mutual Gaze Detection using Pseudo 3D Gaze

Authors: Bardia Doosti, Ching-Hui Chen, Raviteja Vemulapalli, Xuhui Jia, Yukun Zhu, Bradley Green1273-1281

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on three image datasets show that the proposed approach improves the detection performance significantly without additional annotations. This work also introduces a new image dataset that consists of 33.1K pairs of humans annotated with mutual gaze labels in 29.2K images.
Researcher Affiliation Collaboration Bardia Doosti ,1 Ching-Hui Chen , 2 Raviteja Vemulapalli 2, Xuhui Jia 2, Yukun Zhu 2, Bradley Green 2 1 Luddy School of Informatics, Computing, and Engineering, Indiana University Bloomington 2 Google Research bdoosti@indiana.edu, {chuichen,ravitejavemu,xhjia,yukun,brg}@google.com
Pseudocode No The paper describes the proposed approach in narrative text and with a system diagram, but does not include any structured pseudocode or algorithm blocks.
Open Source Code No The paper provides a link for a dataset ('This dataset can be downloaded from https://research.google/tools/datasets/google-openimages-mutual-gaze-dataset/'), but there is no explicit statement or link indicating the availability of the source code for the described methodology.
Open Datasets Yes We introduce a new in-the-wild image dataset that consists of 33.1K pairs of humans annotated with mutual gaze in 29.2K images. This dataset can be downloaded from https://research.google/tools/datasets/google-openimages-mutual-gaze-dataset/. We use three image-based mutual gaze detection datasets in our experiments, namely UCO-LAEO (Marin-Jimenez et al. 2019), AVA-LAEO (Gu et al. 2018) and Open-Images-MG (OI-MG).
Dataset Splits No The paper specifies training and test splits for all datasets (e.g., 'This dataset is divided into 26,410 training pairs and 6,659 test pairs with zero overlap between training and test images' for OI-MG), but does not explicitly detail a separate validation dataset split.
Hardware Specification No The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running its experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers, such as programming languages, libraries, or frameworks used for the experiments.
Experiment Setup Yes In all our experiments, the value of loss function parameter λ (equation (5)) was set to 1.0. All the models were trained using RMSprop optimizer with mini-batches of 128 samples. We used an initial learning rate of 5 10 4 which was gradually reduced by a factor of 0.94 after every 100K steps until convergence. To increase the robustness of the models, during training, we randomly applied horizontal flipping to the images, jittered the head bounding boxes, and adjusted the image intensity and contrast.