Research Projects
Sparse Activation Maps for Interpreting 3D Object Detection
Advisor: Prof. Xiaojun Qi
Proposed Sparse Activation Mapping (SAM) to provide visual explanation for 3D object detection networks, without a need for architecture modification or re-training
Introduced a customized average pooling of convolutional layer weights to maintain the sparsity pattern of point clouds.
Applied SAMs to compare the effectiveness and trustworthy of different volumetric-based 3D object detection models. To the best of our knowledge, this is the first systematic interpretations framework of these methods.
Facial Expression Recognition with Geometric Constrained Attention on Keypoints
Advisor: Prof. Xiaojun Qi
Proposed an end-to-end network with attention on facial keypoints for face expression classification in the wild, then applied geometric constraints on facial keypoints detections to filter out occluded areas.
Outperformed the state-of-the-art face expression classification method on AffectNet dataset, especially robust in challenging occlusions.
Self-attention Point-Voxel Network for 3D Object Detection
Advisor: Prof. Xiaojun Qi, Prof. Ziqi Song
Proposed a robust network for 3D object detection in point cloud, which exploits a novel self-attention algorithm to model the relationship between keypoints from voxel set abstraction.
Utilized a novel computational efficient decoder to fuse the feature maps from multiple receptive fields, which outperforms our baseline model in KITTI dataset.
