Dense Prediction
Overview
Dense prediction studies how models assign meaningful outputs to every pixel, patch, or grid cell, supporting tasks such as segmentation, depth estimation, pose, and weather forecasting. Common approaches combine multi-scale CNN or transformer features with task-specific decoders, while self-supervised pretraining, distillation, and foundation-model adaptation aim to improve transfer across tasks and domains. Evaluation spans benchmarks such as Cityscapes and ADE20K for segmentation, COCO for detection and pose, and KITTI or NYUD-v2 for depth. Current work seeks to retain fine spatial detail while reducing compute and adapting to varied data, with a continuing tension between unified models and specialized systems.
Latest papers
34 papers5 papers in the last four weeks, down 29% on the four weeks before. 0.0% of all new papers.
Adapting Sleep Foundation Models for Dense Sleep Spindle Localization
SAGE: Saliency-Annealed Gated Encoder for Universal Vision Recognition
CODI: Improving Self-Supervised Vision Transformers with Cross Distillation
Supervision-Adaptive Curriculum and Guarded Anti-Curriculum Learning for Dense Prediction
Unlocking Fine-Grained Perception in CLIP via Structurally-Aware Latent Masked Modeling
PixelDense: Dense Prediction as Representation Alignment for Pixel Diffusion
Decoupling Spherical Reasoning from Dense Prediction for 360 Depth Estimation
Scalable In-Domain Self-Supervised Foundation Model for Dense Representation Transfer in High-Resolution Plant Imaging
GTR: Gated Token Recurrence for Efficient Dense Prediction
val2017 with 1.908,ms median batch-one latency under compiled FP16 execution on an RTX