This repository provides a PyTorch re-implementation of Panoptic-DeepLab, a bottom-up approach to panoptic segmentation. It aims to offer a simple, strong, and fast baseline for assigning semantic and instance labels to every pixel in an image, targeting researchers and practitioners in computer vision.

How It Works

Panoptic-DeepLab employs a bottom-up strategy, first detecting object centers and then segmenting instances and semantic regions around these centers. It leverages the DeepLabV3/V3+ architecture with atrous separable convolutions for efficient feature extraction and incorporates a novel approach to combine semantic segmentation and instance prediction, reportedly achieving state-of-the-art results.

Quick Start & Requirements

The project recommends using the Detectron2 implementation. Installation and usage details are available in tools_d2/README.md. Training can be performed on multiple GPUs (e.g., 4x 1080Ti).

Highlighted Details

• Achieves 63.4% PQ on Cityscapes using an HRNet-48 backbone.
• Reproduces COCO results with 35.5 PQ.
• Supports DepthwiseSeparableConv2d (DSConv) for efficiency.
• Offers support for various backbones including ResNet, ResNeXt, and HRNet.

Maintenance & Community

The primary implementation is being deprecated in favor of the Detectron2 version, which is maintained by the author. Contact: Bowen Cheng (bcheng9 AT illinois DOT edu).

Licensing & Compatibility

The repository does not explicitly state a license. It acknowledges utility functions from DeepLab and Detectron2. Compatibility for commercial use or closed-source linking is not specified.

Limitations & Caveats

The author notes that the implementation in this repository will be deprecated and may not reproduce all paper numbers, recommending the Detectron2 version for slightly better results. Post-processing code for deployment is not included and is stated to be a bottleneck.