Six Alibaba CVPR 2022 Papers: Summaries and Open‑Source Releases

CVPR (Conference on Computer Vision and Pattern Recognition) is one of the three top computer‑vision conferences, organized by IEEE. In the latest edition it received 8,161 submissions and accepted 2,067 papers (≈25% acceptance rate).

Alibaba’s DaTaobao technology team had six papers accepted at CVPR 2022, and all code and models are publicly released. The papers cover: • Video restoration with a two‑stage network that combines multi‑frame RNN and single‑frame Transformer, using transfer learning and progressive training. • Real‑time image enhancement via a learned non‑uniform 3‑D color lookup table, achieving the highest PSNR and fast 4K video processing. • GEN‑VLKT for human‑object interaction detection, transferring knowledge from large‑scale pre‑trained models and setting SOTA on regular and zero‑shot tasks. • Efficient indirect illumination modeling for inverse rendering, decoupling illumination from BRDF and environment light. • NeRF‑Editing, a method that enables free geometry editing of neural radiance fields. • A robust NeRF reconstruction approach that improves view synthesis when test viewpoints differ greatly from training views, now used in Alibaba’s Object Drawer project.

Readers are invited to watch the live stream for detailed paper walkthroughs and to reserve a spot for deeper insights.