ECCV 2026 has announced this year’s best paper, best paper nominees, test-of-time awards, and other honors. ECCV, the European Conference on Computer Vision, is one of the field’s three top computer vision conferences. The 2026 edition took place in Malmo, Sweden, from Sept. 8 to 12.
According to figures cited by the source article, ECCV 2026 received 10,473 valid submissions and accepted 2,834 papers, for an acceptance rate of 27.1%. Of those, 163 papers were chosen for oral presentation, or 1.6%. The article also says authors from China have become a major force at ECCV.
Best paper goes to HKTex work from Imperial College London
The conference’s best paper award went to “Heat Kernel Textures -- the Geodesic Gaussians That Do Not Splat,” from Imperial College London. The two co-first authors are postdoctoral researcher Simone Foti and PhD student Caner Korkmaz. Stefanos Zafeiriou and Tolga Birdal are also on the author list.
The award citation praised the paper for “its simple and novel contribution to surface texture representation.” It says the method swaps out standard UV mapping and Gaussian splatting for heat-diffusion-based geodesic Gaussians on non-Euclidean manifolds. The committee called the method elegant and said its promise goes beyond this paper alone.
Paper link: https://eccv.ecva.net/virtual/2026/poster/3652
The paper introduces Heat Kernel Textures, or HKTex, as a new representation for 3D texture. The basic idea borrows from 3D Gaussian Splatting, which uses large numbers of Gaussian-like primitives to encode visual information. But HKTex puts heat-kernel primitives directly on a mesh surface, not in 3D Euclidean space.
Most traditional 3D models still depend on UV mapping. That means flattening a 3D surface into a 2D image and storing color there. The article says this old setup comes with familiar headaches: distortion during unwrapping, seams, uneven resolution from region to region, and big empty chunks in UV maps that waste memory and storage. And for complex models, seams can force duplicated vertices too.
HKTex drops the UV map entirely. The article describes it as scattering thousands of colored spots across a model surface, with each one acting like a Gaussian that diffuses along the surface through geodesic distance. In mathematical terms, the method uses anisotropic heat kernels, described as the surface-space version of a Gaussian. So the texture primitive can bend with the surface naturally, instead of drifting off the model the way standard 3D Gaussians sometimes do.
Even though the method is inspired by Gaussian Splatting, it does not use splatting for rendering. Instead, rays first hit the mesh, and then the renderer dynamically computes the color contribution from nearby heat kernels at the hit point. That makes it compatible with differentiable physically based rendering and ray tracing pipelines.
These heat kernels are trainable. During optimization, the system can update their position, orientation, shape, scale, and color. But their motion stays strictly locked to the mesh surface. It also runs a densification-style process, similar to 3DGS, adding more kernels in texture-heavy areas and fewer in simpler ones.
The paper gives two use cases. One is compressing an existing UV texture into HKTex. The other is recovering object-surface texture directly from multi-view photographs through inverse rendering, without creating UVs first.
In texture-fitting experiments on 313 Objaverse models, the authors said HKTex sharply reduced texture storage while keeping image quality high. The source article says the appendix reports roughly an order-of-magnitude drop in texture storage compared with the original ground-truth UV texture. In texture recovery from multi-view images, HKTex used an average of about 78.7 KB. It also outperformed the neural texture and vertex texture baselines used by the authors on PSNR, LPIPS, SSIM, MS-SSIM, and MSE.
Two papers named best paper nominees
ECCV 2026 also picked two best paper nominees, with authors from Meta, Stony Brook University, and other institutions.
LSRM from Meta Reality Labs Research
One nominee was “LSRM: High-fidelty Object-centric Rsconstruction via Scaled Context Windows,” by Zhengqin Li, Cheng Zhang, Jakob Engel, and Zhao Dong.
Paper link: https://eccv.ecva.net/virtual/2026/poster/4770
The award citation said the paper stood out for very high quality and strong engineering execution. It pushes object-centric 3D reconstruction forward by extending context window length in a Transformer architecture, allowing it to capture long-range cross-view dependencies and produce higher fidelity.
This work comes out of Meta Reality Labs Research and introduces LSRM, short for Large Sparse Reconstruction Model. The paper asks a pretty direct question: if you make the Transformer context window much bigger, can you get better feed-forward 3D reconstruction quality?
The starting point is simple. Recent object-centric feed-forward reconstruction methods are fast and reliable, but they still lag behind per-scene optimized dense-view methods when it comes to fine texture and appearance recovery. The authors argue that a lot of that gap comes down to token budget. In their view, if the number of object tokens and image tokens in computation is expanded sharply, a feed-forward system can also achieve high-fidelity 3D object reconstruction and inverse rendering.
To make that scale practical, LSRM brings native sparse attention into 3D reconstruction and adds three changes:
- a coarse-to-fine pipeline that predicts sparse high-resolution residuals and puts compute on information-rich regions;
- a 3D-aware spatial routing mechanism that uses explicit geometric distance, instead of ordinary attention scores, to build 2D-3D correspondence;
- a sequence-parallel strategy for block-structured data, paired with an All-gather-KV communication protocol to balance dynamic and sparse workloads across multiple GPUs.
Taken together, the paper says LSRM can process 20 times as many object tokens as the previous state of the art, and more than twice as many image tokens. On standard novel-view synthesis benchmarks, it posted a PSNR gain of more than 2.4 dB over the current state of the art and reduced LPIPS by more than 40%. And when the method was extended to inverse rendering, the paper reports consistent gains in texture and geometric detail on common benchmarks, with LPIPS matching or beating dense-view optimization methods.
Poppy targets stubborn failures in monocular normal estimation
The other nominee was “Poppy: Polarization-Based Plug-and-Play Guidance for Enhancing Surface Normal Estimation.”
Paper link: https://eccv.ecva.net/virtual/2026/oral/6004
Project page: https://irnkim.github.io/poppy/
The author team includes Irene Kim, Sai Tanmay Reddy Chakkera, Alexandros Graikos, Dimitris Samaras, and Akshat Dave.
The citation said the paper made an important contribution to test-time adaptation across multiple pure-RGB backbone networks. With just a single polarization capture, it resolves ambiguities, especially in hard non-polarized regions and scenes with complex specular reflections.
The paper comes from Stony Brook University and goes after a failure case that has refused to go away in monocular surface normal estimation. Models trained on large RGB-to-normal datasets often struggle on reflective, textureless, and dark surfaces. Not surprising. Those areas just do not give the network many visual cues to work with.
The authors bring in polarization as a physical complement. Polarization signals carry surface orientation information and do not depend on texture or albedo, which makes them a natural match for the weak points of RGB-based methods. The article says existing polarization approaches often need multi-view capture or dedicated training data, which limits how well they generalize.
Poppy is built as a training-free framework. At test time, it uses polarization measurements from a single capture to optimize the normal predictions of any frozen RGB backbone. The backbone weights never change. Poppy instead optimizes per-pixel offsets while jointly learning a reflectance decomposition.
Its core piece is a differentiable rendering layer. That layer turns corrected normals into predicted polarization values, then compares them with the observed polarization signals, using the mismatch as a penalty to drive optimization. In the paper’s setup, physical observation serves as supervision at test time, not as a source of training data. That is why the framework can be attached to existing backbones.
Evaluation covered seven benchmarks and three backbone classes: diffusion models, flow models, and feed-forward models. The results show a 23% to 26% reduction in mean angular error on synthetic data and a 6% to 16% reduction on real data. From that, the authors said test-time polarization cues can improve normal quality on difficult surfaces without retraining.
Three test-of-time awards, including a paper co-authored by Fei-Fei Li
Along with the three highlighted papers, ECCV also released a list of 10 finalist papers. This year’s test-of-time awards went to three papers published a decade ago.
The first was “SSD: Single Shot MultiBox Detector,” available at https://arxiv.org/abs/1512.02325, by Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C. Berg. The citation said it helped push object detection into the deep learning era, alongside the R-CNN and YOLO model families.
The second was “Perceptual Losses for Real-time Style Transfer and Super-resolution,” available at https://arxiv.org/abs/1603.08155, by Justin Johnson, Alexandre Alahi, and Fei-Fei Li. The citation said it introduced perceptual loss functions based on pretrained neural features, a method now used so widely that people almost forget where it came from.
The third was “Learning Without Forgetting,” available at https://arxiv.org/abs/1606.09282, by Zhizhong Li and Derek Hoiem. The citation said it pioneered the idea of continual learning, which is still an open challenge and still a busy area of vision AI research.
ECVA also released details on PhD and young researcher awards
The European Computer Vision Association also published details on its PhD Award. The award is meant to recognize and encourage outstanding research achievements in computer vision at the doctoral dissertation stage. Each winner receives a 2,500 euro prize. Two awards are given each year and presented at the next ECCV conference.
The ECVA Young Researcher Award is meant to recognize and encourage outstanding work by early-career researchers in computer vision. Each winner receives 5,000 euros. One award is given each year and presented at the next ECCV conference.
The source article was published by the WeChat public account “Heart of the Machine” (ID: almosthuman2014) and credited to “Heart of the Machine, focused on AI.”