Setup

The following guide shows how to setup an environment in order to train or evaluate Deep-Learning models using the GOOSE Dataset.

Installation

In this guide, we use the SuperGradients [1] to conveniently load and evaluate different models. You can use either conda or venv to create a new virtual environment with the following dependencies:

name: sg_env
channels:
  - pytorch
  - nvidia
dependencies:
  - python==3.9
  - pytorch==1.13.1
  - torchvision==0.14.1
  - pytorch-cuda==11.7
  - pip
  - pip:
    - super_gradients==3.2.0
    - matplotlib
    - pillow
    - numpy
    - torchmetrics==0.8.0

Save the requirements as file and create the environment (e.g. using conda env create -f env.yaml). Then activate the environment (e.g. using conda activate sg_env) and you are good to go.

Download Dataset

You can directly download the preconfigured and zipped raw image and ground truth files:

• training split (22.5 GB)
• validation split (2.9 GB)
• test split (3.4 GB, raw images only)

The direct download for annotated pointclouds will follow soon. The GOOSE-DB can be used to download entire sequences with all available metadata. After registering, you can select sequences with specific annotation for download.

Pretrained Weights

We provide pretrained weights for some network architectures both for 2D and 3D semantic segmentation.

2D Image Segmentation

We mainly evaluate our dataset on ppliteseg and ddrnet networks, which both have a good tradeoff between realtime capabilities and quality.

• PP-LiteSeg [2] uses an encoder-decoder structure with a lightweight attention-based fusion model.
• DDRNet [3] uses a typical two-stream architecture which fuses both branches at different depths within the network.

Model	Model Name [Download]	Dataset	Resolution	# Classes	mIoU / %
PP-LiteSeg	ppliteseg_category_512	GOOSE-2D	512x512	12	67.21
PP-LiteSeg	ppliteseg_class_512	GOOSE-2D	512x512	64	45.09
DDRNet	ddrnet_category_512	GOOSE-2D	512x512	12	70.23
DDRNet	ddrnet_class_512	GOOSE-2D	512x512	64	46.53

3D Pointcloud Segmentation

References

1. Aharon et al. "Super-Gradients" https://zenodo.org/record/7789328 (2021)
2. Peng et al. "PP-LiteSeg: A Superior Real-Time Semantic Segmentation Model" https://arxiv.org/abs/2204.02681 (2022)
3. Pan et al. "Deep Dual-Resolution Networks for Real-Time and Accurate Semantic Segmentation of Traffic Scenes" in IEEE Trans. Intell. Transp. Syst. (2022)