Shape Vision Lab UChile

We wish him continued success in his scientific journey! Full post ➡️: lnkd.in/d2BVUc8h #AI #ComputerVision #3DReconstruction #CulturalHeritage #Research #ShapeVisionLab

Congratulations to our student Cristián Llull for successfully passing his doctoral qualification exam! This achievement marks an important first step on his path earning a Ph.D. We are confident he will continue to demonstrate dedication, research excellence and academic rigor.

SHREC 2026: reconstruct high-frequency geometry from 90 views (COLMAP poses). Dataset out now. Registration → cllull@dcc.uchile.cl. Submissions due Apr 3, 2026. Details: shapevision.dcc.uchile.cl/cllull-shrec20… #ComputerVision #3DReconstruction #Photogrammetry

What if we could model vision like a wave moving through space? Researchers from Peking & Tsinghua Universities present WaveFormer. They treat image features as signals governed by a wave equation, explicitly controlling how low-to-high frequency details evolve across network layers. This new Wave Propagation Operator outperforms standard Vision Transformers in image classification, detection, and segmentation, achieving up to 1.6x higher throughput with 30% fewer computations. WaveFormer: Frequency-Time Decoupled Vision Modeling with Wave Equation Paper: arxiv.org/abs/2601.08602 Code: github.com/ZishanShu/Wave… Our report: mp.weixin.qq.com/s/xFoj94IIG4xj… 📬 #PapersAccepted by Jiqizhixin

DeepSeek just dropped a banger paper to wrap up 2025 "mHC: Manifold-Constrained Hyper-Connections" Hyper-Connections turn the single residual “highway” in transformers into n parallel lanes, and each layer learns how to shuffle and share signal between lanes. But if each layer can arbitrarily amplify or shrink lanes, the product of those shuffles across depth makes signals/gradients blow up or fade out. So they force each shuffle to be mass-conserving: a doubly stochastic matrix (nonnegative, every row/column sums to 1). Each layer can only redistribute signal across lanes, not create or destroy it, so the deep skip-path stays stable while features still mix! with n=4 it adds ~6.7% training time, but cuts final loss by ~0.02, and keeps worst-case backward gain ~1.6 (vs ~3000 without the constraint), with consistent benchmark wins across the board