zhanghang1989/ResNeSt

ResNeSt: Split-Attention Networks

What it solves

ResNeSt is a variant of the ResNet architecture designed to improve the performance of computer vision tasks. It specifically targets the limitations of standard ResNets by introducing a split-attention mechanism to better capture cross-channel dependencies and scale-related features.

How it works

The project implements "Split-Attention Networks," which modify the standard residual block to allow the model to learn a more flexible way of aggregating information across different groups of channels. This architecture can be used as a backbone for various downstream tasks such as object detection, instance segmentation, and semantic segmentation.

Who it’s for

This is for computer vision researchers and developers who need a high-performance backbone network for image classification, object detection (e.g., via Mask R-CNN or Cascade R-CNN), and semantic segmentation (e.g., via DeepLabV3).

Highlights

  • Multi-framework support: Provides implementations for PyTorch and MXNet Gluon, with 3rd party versions for TensorFlow, Caffe, and JAX.
  • Downstream boost: Significantly improves performance for models like Mask R-CNN, Cascade R-CNN, and DeepLabV3.
  • Pretrained models: Offers a variety of pretrained weights for different model sizes (ResNeSt-50, 101, 200, and 269).
  • Integration: Includes wrappers for Detectron2 and integration with MMDetection.

Related

  • Project
    deepdoctection/deepdoctection

    A Python library for document understanding that orchestrates layout analysis, OCR, and token classification to extract structured data from PDFs and scans.

  • Project
    oculix-org/SikuliX1

    A computer vision-based automation tool that identifies and interacts with screen elements using image recognition to simulate mouse and keyboard input.

  • Project
    hbb1/2d-gaussian-splatting

    A 2D Gaussian Splatting implementation that uses oriented disks (surfels) to create geometrically accurate radiance fields and high-quality 3D meshes.

  • Project
    Peterande/D-FINE

    D‑FINE is a real‑time object‑detection framework that redefines DETR’s box regression as a fine‑grained distribution refinement and adds a self‑distillation module. It offers several model sizes (‑N to ‑X) that achieve state‑of‑the‑art COCO/AP scores (42.8 %–55.8 %) while running at 70‑470 FPS on a T4 GPU, with no extra inference cost. The repo provides pretrained checkpoints, config files, and full training/evaluation scripts for COCO, Objects365, and custom COCO‑style datasets.