Highlights
- Spatiotemporal Point Cloud Processing
- The novel problem of sequential scene flow estimation (SSFE); baselines and dataset benchmark
- The novel recurrent set-to-set cost volume layer for spatiotemporal point cloud processing
- The first scene flow estimation (SFE) framework via recovering motion from the alignment of PDFs
- The novel Cauchy-Schwarz (CS) divergence-based self-supervised objective for SFE
- The first LLE-inspired algorithm for unsupervised shape correspondence by representing maps between pairs of shapes as locally linear transformations. It is different from the classic functional map framework
- The first work that addresses sparse computation frameworks on multiple GPUs with pipeline model parallelism.
- Scene Text Understanding
- The first deep learning framework that regards scene text recognition (STR) as a sequence labelling problem
- The popular Connectionist Text Proposal Network (CTPN) framework detecting a text line in a sequence of fine-scale text proposal
- The Single Shot Text Detector that introduces regional attention for text detection
- The Guided CNN for faster inference on popular text detectors
- The first adversarial attack approach---adaptive attack---on STR systems to speed up the generation of adversarial examples
- More research can be found in my google scholar
Demo & Code
The novel problem of SSFE from point cloud sequences; baselines and benchmark
The first LLE-inspired algorithm for unsupervised shape correspondence by representing maps between pairs of shapes as locally linear transformations
A novel self-supervised CS objective; The first SFE framework that converts SFE into the problem of recovering motion from the alignment of PDFs
Click for the web demo
Demo & Code for Single Shot Text Detector; state-of-the-art text detector
Demo & Code for Connectionist Text Proposal Network; state-of-the-art text line detector
Click for the web demo
Deployed at Gainesville for improving pedestrain safety
Semi-annotation tool and infrastructure dataset; Real-world deployment.
A new large-scale egocentric video benchmark recorded by takeaway riders about their daily work
Unsupervised object-centric learning
an efficient and asynchronous parallelism approach for handling 3D point clouds with distributed training