Computer Vision News
The latest Computer Vision coverage curated for game developers.
arXiv cs.GR details a pipeline that turns figurative paintings into multiple plausible 3D interpretations instead of forcing one reconstruction. It samples camera-orbit videos, rebuilds them with 3D Gaussian Splatting, then …
arXiv cs.GR details SplashSplat, a new method for reconstructing splashing liquids from real multi-view video. The team pairs a 20-scene benchmark with seven synchronized 4K cameras at 60 fps and …
arXiv cs.GR details a model that inserts 3D humans into partial graphic designs, aiming to make layouts feel more deliberate and visually coherent. The system predicts pose, framing, and placement …
arXiv cs.GR details CADSplat, a sparse-view 3D Gaussian splatting pipeline that uses CAD models to reconstruct photorealistic digital twins from fewer than 15 posed images. For teams building AR, inspection, …
Researchers have pushed non-line-of-sight imaging beyond the usual three-bounce assumption, using higher-order light paths to reconstruct hidden objects around one or even two corners. The technique combines ultrafast laser scanning …
PhysStream pushes video generation toward physically meaningful control, letting creators steer motion mid-generation instead of locking in a full schedule up front. The model uses structured scene memory and sparse …
A new hand-gesture method targets frame-accurate synchronization instead of loose classification, using a single projected winding-order signal to detect an open-hand flip. It needs no training data, calibration, or classifier, …
SyntheticDoc brings 1,000,000 procedurally generated, high-resolution training images to document unwarping and illumination correction. The dataset pairs each sample with pixel-perfect UV, normal, albedo, and shading maps, aiming to replace …
Researchers have introduced Hi-SPAD, a video-rate hyperspectral imaging and inference system built around single-photon cameras. The work targets fast spectral capture that could improve material classification, lighting analysis, and other …
A new HDR reconstruction method extends highlights one exposure step at a time instead of solving the full scene in one pass. The recurrent approach is trained on widely available …
RealSimLoop uses vision feedback and differentiable reduced-order simulation to adapt deformable-object models online, aiming for quasi-real-time real-to-sim calibration. For teams working on soft-body physics, robotics-style perception, or VFX-heavy interactions, the …
A new vision-language system can turn origami demonstration videos into executable folding programs, using geometry checks and physical simulation to keep the sequence plausible. For game teams, the interesting part …
NeuSOGA turns geometric observations into editable symbolic math, aiming to replace opaque latent encodings with explicit representations. The framework combines topology-guided discovery, Segment Anything-based perception, multi-scale abstraction, and Implicit Area …
Deep-learning depth estimation for endoscopic surgery is getting a broad technical map, with a new survey organizing the field around data, methods, and clinical use. For developers working in medical …
A new reconstruction method turns curved capacitive textiles into a 3D shape sensor, letting robots infer nearby geometry from proximity fields instead of flat RGB or depth inputs. Proximity3D aggregates …
Khronos has released OpenVX 1.3.2, tightening error reporting and consistency for vision pipelines. The update adds VX_ERROR_TIMEOUT and VX_ERROR_GRAPH_NOT_VERIFIED, expands image support with RGBA and U1 handling, and sets up …
A new copper-oxidation pipeline forecasts how a fixed surface will look 10 accelerated units ahead, then turns that prediction into PBR maps for albedo, normal, roughness, and metallic. The practical …
A new framework cuts the cost of building urban segmentation training data by turning rough synthetic scenes into target-aligned images. It adapts an off-the-shelf diffusion model with only imperfect pseudo-labels, …
A new intrinsic decomposition method tackles a long-standing misattribution problem in point-based 3D scene reconstruction. Intrinsic PAPR replaces translucent volume aggregation with proximity-aware point rendering so each surface point can …
EditStream folds text-to-video, image-to-video, video-to-video, editing propagation, reference-guided edits, and camera pose changes into one DiT-based system. The big shift for developers is its few-step autoregressive streaming setup, aimed at …