- ZMO.AI (startup), China
- Research Intern, Jan. 2021 - Apr. 2021, Shenzhen, China
- Realistic and real-time background replacement for on-model images, used for content generation for e-commerce brands.
- Hisense Group Holdings Co., Ltd., China
- Research Intern, Dec. 2020 - May 2021, Qingdao, CHina
- High-quality, high-resolution, and high-fps harmonization on Hisense’s virtual household devices; held three relevant patents.
- University of Texas at Austin, USA
Remote Visiting Student, ECE department, Jun. 2021 - Dec. 2021
- Object-Centric Image Unfolding: On-going work on a new challenging task.
- Shanghai Jiao Tong University, Shanghai, China
Research Assistant, CS department, Dec. 2018 - Mar. 2022
- High-Resolution Image Harmonization [code]: Proposed the first high-resolution image harmonization network with collaborative dual transformations: high-resolution traditional color transformation and low-resolution deep pixel transformation; achieved ~50% improvement and saved ~65% time and computational resources.
- Cross-Domain Image Harmonization [code]: Proposed the first cross-domain image harmonization network using the mixture of rendered images and real images; achieved ~35% improvement on novel categories.
- Rendered Human Harmonization Dataset Construction: Released the first large-scale rendered image harmonization dataset RHHarmony with 135k image pairs; mitigated labor-intensive extension and supplemented real datasets.
- Deep Video Harmonization [code]: Leveraged color mapping consistency to lift the burden of establishing spatial correspondence; achieved ~15% improvement than the strongest baseline with higher temporal consistency.
- Video Harmonization Dataset Construction: Released the first public, large-scale, and high-quality video harmonization dataset HYouTube which contains 3194 pairs of synthetic composite videos and real videos.
- Background-Guided Image Harmonization [code]: Reformulated image harmonization as background-guided domain translation to explicitly utilize background domain to guide the foreground harmonization; enabled inharmony level assessment of composite images; achieved new state-of-the-art: ~20% improvements than the strongest baseline.
- Image Harmonization via Domain Verification [code]: Proposed novel domain verification discriminator to pull close the foreground domain and background domain; achieved state-of-the-art performance: ~30% improvements than the strongest baseline; provided the first benchmark in image harmonization field.
- Image Harmonization Dataset Construction: Released the first large-scale image harmonization dataset iHarmony4 with 4 sub-datasets (HCOCO, HAdobe5k, HFlickr, and Hday2night) and 73146 pairs of high-quality images.