Yitong Jiang

Yitong Jiang, Hongjun Wang, Collin McCarthy et al.

Vision foundation models are bottlenecked by the quadratic cost of self-attention, which limits usable resolution and increases the cost of large-scale pretraining. Subquadratic alternatives such as linear attention and state-space models reduce this cost, but often serialize images into 1D token streams and weaken the 2D spatial structure important for vision. Generalized Spatial Propagation Networks (GSPN) instead propagate context directly on the 2D grid through line-scan recurrences, achieving near-linear complexity without positional embeddings, but have seen little use as foundation-scale encoders. We present C-GSPN, a foundation-scale vision encoder based on 2D spatial propagation. C-GSPN makes the operator practical through three improvements: (1) a fast GSPN CUDA kernel that fuses per-step launches into a single warp-specialized implementation with shared-memory tiling, coalesced access, and a compact multi-channel propagation, reaching over 90% of peak memory bandwidth and running up to 40--52x faster than the original GSPN implementation; (2) a compressed latent-space propagation block with fused normalization, which turns kernel-level speed into block- and model-level efficiency; and (3) a two-stage cross-operator distillation recipe that trains the new architecture from an attention teacher without the cost of from-scratch foundation-scale training. Distilled with 600M image-text pairs, C-GSPN matches an isomorphic ViT baseline with 15% fewer parameters, improves ADE20K segmentation by +2.1%, transfers to high resolution with a fraction of the data needed from scratch, and delivers a 4x end-to-end block speedup at 2K with single-pass, tiling-free inference.

Yitong Jiang, Inchang Choi, Jun Jiang et al.

We propose a novel high dynamic range (HDR) video reconstruction method with new tri-exposure quad-bayer sensors. Thanks to the larger number of exposure sets and their spatially uniform deployment over a frame, they are more robust to noise and spatial artifacts than previous spatially varying exposure (SVE) HDR video methods. Nonetheless, the motion blur from longer exposures, the noise from short exposures, and inherent spatial artifacts of the SVE methods remain huge obstacles. Additionally, temporal coherence must be taken into account for the stability of video reconstruction. To tackle these challenges, we introduce a novel network architecture that divides-and-conquers these problems. In order to better adapt the network to the large dynamic range, we also propose LDR-reconstruction loss that takes equal contributions from both the highlighted and the shaded pixels of HDR frames. Through a series of comparisons and ablation studies, we show that the tri-exposure quad-bayer with our solution is more optimal to capture than previous reconstruction methods, particularly for the scenes with larger dynamic range and objects with motion.

Xiaowei Hu, Tianyu Wang, Chi-Wing Fu et al.

Shadow detection in general photos is a nontrivial problem, due to the complexity of the real world. Though recent shadow detectors have already achieved remarkable performance on various benchmark data, their performance is still limited for general real-world situations. In this work, we collected shadow images for multiple scenarios and compiled a new dataset of 10,500 shadow images, each with labeled ground-truth mask, for supporting shadow detection in the complex world. Our dataset covers a rich variety of scene categories, with diverse shadow sizes, locations, contrasts, and types. Further, we comprehensively analyze the complexity of the dataset, present a fast shadow detection network with a detail enhancement module to harvest shadow details, and demonstrate the effectiveness of our method to detect shadows in general situations.