Computer Vision
Depth-Aware Distillation Enhances Visual Place Recognition in Forests
This paper introduces a lightweight, depth-aware distillation framework to enhance visual place recognition in challenging forest environments. By integrating geometric cues from depth data into a DINOv2-based model without altering its descriptor space, the approach improves robustness against appe…
Triangle Splatting Bridges Differentiable Rendering and Explicit Geometry for Real-Time SLAM
Triangle Splatting SLAM introduces the first dense SLAM system that uses differentiable triangles as the 3D map representation, enabling both tracking and mapping through online differentiable rendering. Unlike 3D Gaussian Splatting which excels at novel-view synthesis, this approach maintains expli…
Geometry-Aware Vision Networks Improve Precision Guidance for Robot-Assisted Endoscopic Surgery
GeoCFNet addresses visual guidance challenges in robot-assisted endoscopic submucosal dissection by modeling dissection regions as geometry-aware confidence fields rather than discrete segmentation masks. The approach integrates pretrained vision foundation models (DINOv3) with specialized spatial r…
BitResEdit: Leveraging Native Model Structures for Enhanced VAR Image Editing
BitResEdit introduces a novel training-free image editing framework for bitwise-residual Visual Autoregressive (VAR) models. It uniquely leverages the per-bit prediction head (BitEdit) and multi-scale residual code field (ResEdit) native to these models. BitEdit guides edits via source-negative guid…
Calibrating Multi-Camera Fusion with Component-Wise Error Analysis for Improved Indoor Localization
This work introduces a measurement-calibrated fusion approach for vision-based indoor localization that explicitly quantifies errors from single-camera components like homography calibration, human detection, and motion tracking. While not significantly improving absolute accuracy over standard fusi…
Unified, Lightweight Architecture for Demosaicing Pixel-Bin Image Sensors with Learning-Free CFA Identification
This paper introduces a unified, modular, and lightweight architecture for demosaicing pixel-bin image sensors, addressing the challenges posed by their increased inter-color separation and the limitations of existing CFA-specific deep learning models. The architecture aims to enhance image quality …
DynaTok: Image-Free 4D Reconstruction from Partial Point Clouds Using Transformers
DynaTok is a novel point-based framework designed for 4D reconstruction from incomplete and unordered partial point cloud sequences, overcoming limitations of existing image-based and object-centric methods. It employs a Transformer-based spatiotemporal encoder to aggregate spatio-temporal informati…
MSUE: Multi-Expert Soccer VQA System Achieves 0.95 Accuracy using VLM and LLM Routing
This paper introduces MSUE, a multi-expert question-answering system for the SoccerNet VQA Challenge. MSUE utilizes a VLM-driven pipeline for data synthesis and an LLM to route questions to specialized experts: Gemini3-Flash for text, a fine-tuned Qwen3-VL for visual/video data, and an external know…
YOLOv12 Automates Network Cable Wire Sequencing Verification with High Accuracy
This paper introduces a YOLOv12-based object detection system for verifying the correct color sequence of wires in network cables on a production line. The system achieved high accuracy, with precision around 98% and mean accuracy of 95%, significantly outperforming traditional manual inspection met…
Lighting-Aware Representation Learning for Improved Visual Understanding
This research introduces a lighting-aware representation learning framework that treats illumination variation as an explicit training signal, diverging from conventional invariance-focused data augmentation. By extending contrastive learning with an auxiliary objective for illumination-dependent va…
ETCH-X: A Robust and Expressive Human Body Fitting Method for Clothed 3D Scans
ETCH-X is a novel human body fitting method designed to improve both the expressiveness and robustness of fitting parametric body models like SMPL-X to 3D point clouds of clothed humans. It achieves this through a tightness-aware fitting paradigm that filters out clothing dynamics, utilizes implicit…
Elastic Test-Time Training for Scalable 4D Reconstruction
The proposed Fast Spatial Memory (FSM) architecture leverages Elastic Test-Time Training to enable multi-chunk adaptation for 4D reconstruction. By incorporating a Fisher-weighted elastic prior and an exponential moving average anchor state, the system mitigates catastrophic forgetting and overfitti…