D2Det is an official implementation of a CVPR2020 paper offering a novel two-stage object detection and instance segmentation method. It targets researchers and practitioners in computer vision seeking improved localization and classification accuracy, particularly for challenging datasets. The core benefit is enhanced precision through unique regression and pooling strategies.

How It Works

D2Det introduces two key innovations: dense local regression and discriminative RoI pooling. Dense local regression predicts multiple dense box offsets for object proposals, moving beyond fixed keypoints to enable more precise localization. This is further refined by a binary overlap prediction strategy to mitigate background influence. Discriminative RoI pooling adaptively samples and weights sub-regions of a proposal to extract more informative features for accurate classification.

Quick Start & Requirements

• Install: Follow INSTALL.md from the mmdetection toolbox.
• Prerequisites: PyTorch 1.1.0, CUDA 9.0/10.0, MMCV 0.4.3.
• Usage: Training and inference commands are provided using tools/train.py, tools/dist_train.sh, tools/test.py, and tools/dist_test.sh. Demo script available at ./demo/D2Det_demo.py.
• Configs: Configuration files are located in configs/D2Det/. Refer to GETTING_STARTED.md for more details.

Highlighted Details

• Achieves 47.5 box AP on COCO test-dev with ResNet101-DCN backbone (single-scale).
• Provides models for both object detection and instance segmentation.
• Supports large vocabulary datasets LVIS and Objects365 via mmdetection v2.1.0.
• Based on the mmdetection and Grid R-CNN plus open-source toolboxes.

Maintenance & Community

The project is an official implementation of a CVPR2020 paper. No specific community channels or active maintenance signals are provided in the README.

Licensing & Compatibility

The project is released under a license that permits academic use, as it's based on mmdetection and Grid R-CNN plus. Specific licensing details for commercial use are not explicitly stated in the README.

Limitations & Caveats

The provided installation requirements (PyTorch 1.1.0, MMCV 0.4.3) are outdated, potentially requiring significant effort to adapt to current library versions. The project's reliance on older dependencies may pose compatibility challenges.