Get Started
Home
Topics
Search
Library
Research questionHow can vision systems skip unneeded image regions or tokens without hurting accuracy?Processing every region or token can spend limited compute on irrelevant image content. Keeping fewer inputs saves resources only if the information needed for the task survives.
AI
Computer Vision
Evaluation & Benchmarks
Image & Video Processing
Inference Optimization
Machine Learning
Multimodal Models
Reinforcement Learning
Research Paper
Small / On-device Models
Technology
Latest papersRecent research connected to this question, newest first.LookThere! Sparse Vision by Reinforced SelectionThe source presents a jointly trained selector and representation extractor and evaluates sparse processing across recognition, segmentation, zero-shot classification, and counting settings. Evidence is limited to the tested vision-transformer models, tasks, image domains, and sparsity levels.research paper · Sep 4, 2026DenseScout: Algorithm-System Co-design for Budgeted Tiny Object Selection on Edge PlatformsThe source studies a lightweight frontend that ranks patch centers and sends selected regions to a downstream detector, together with transport-aware execution. Its evidence comes from VisDrone and DOTA evaluations and profiling on Jetson Orin NX and RK3588, so it does not establish behavior beyond those tasks and deployment settings.research paper · Sep 4, 2026Efficient Semantic Understanding from Digital FoveationThe evidence comes from an active-vision pipeline using saliency-driven fixation selection, high-resolution foveal observations, low-resolution context, semantic accumulation, and adaptive computation. Results are reported on ADE20K-Object using object-level Top-1 and Top-3 accuracy and scene-level object recall against dense-prediction baselines; broader generalization beyond this benchmark and setup is not established.research paper · Sep 3, 2026Who Speaks for the Pruned? Visual Token Pruning as Coverage OptimizationApplies to visual token pruning across multiple vision-language model architectures and compression rates. The source reports evidence for a training-free approach using projected visual-token representations and a lightweight first-layer attention probe, with gains evaluated primarily through accuracy under compression.research paper · Sep 2, 2026
Related questions
How can high-resolution diffusion Transformers prune tokens without sacrificing image fidelity or predictable compute?How can Vision Transformers avoid LayerNorm’s global reduction on edge accelerators without expensive retraining?How can visual tool pipelines adapt when dense, occluded scenes or domain shift defeat fixed orchestration?How can few-step visual generators preserve preference-aligned quality without being capped by multi-step teacher distillation?