PowerBEV - Woven Dataset

• Contributors : Aman Sharma , Vyshnav Achuthan , Neha Madhekar , Rishikesh Jadhav , Wu Xiyang

Contents

• PowerBEV2
  • Setup
  • Dataset
  • Pre-trained models
  • Training
  • Prediction
    • Evaluation
    • Visualisation
  • Credits
  • Citation

Setup

Create the conda environment by running

conda env create -f environment.yml

Dataset

• Download the full Toyota Woven Planet Perception datset, which includes the Mini dataset and the Train and Test dataset.
• Extract the tar files to a directory named lyft2/ . The files should be organized in the following structure:

  lyft2/
  ├──── train/
  │     ├──── maps/
  │     ├──── images/
  │     ├──── train_lidar/
  │     └──── train_data/
  

Pre-trained models (Comparision)

The config file can be found in powerbev/configs . You can download the pre-trained models which are finetuned for nuscenes dataset using the below links:

Weights	Dataset	BEV Size	IoU	VPQ
PowerBEV_long.ckpt	NuScenes	100m x 100m (50cm res.)	39.3	33.8
PowerBEV_short.ckpt	NuScenes	30m x 30m (15cm res.)	62.5	55.5
PowerBEV_static_long.ckpt	None	100m x 100m (50cm res.)	39.3	33.8
PowerBEV_static_short.ckpt	None	30m x 30m (15cm res.)	62.5	55.5

Training

To train the model from scratch on Woven, run

python train.py --config powerbev/configs/powerbev.yml

and make sure you make the respective changes on the config.yaml file inside configs folder.

For running on pretrained weights

python train.py --config powerbev/configs/powerbev.yml \
                PRETRAINED.LOAD_WEIGHTS True \
                PRETRAINED.PATH $YOUR_PRETRAINED_STATIC_WEIGHTS_PATH

Prediction

Evaluation

To run from the model which was trained from scratch just search for the tensorboard log file which will have the ckpt file and add that ckpt file path as your pretrained weights path.

python test.py --config powerbev/configs/powerbev.yml \
                PRETRAINED.LOAD_WEIGHTS True \
                PRETRAINED.PATH $YOUR_PRETRAINED_WEIGHTS_PATH

Visualisation

To run from the model which was trained from scratch just search for the tensorboard log file which will have the ckpt file and add that ckpt file path as your pretrained weights path.

python visualise.py --config powerbev/configs/powerbev.yml \
                PRETRAINED.LOAD_WEIGHTS True \
                PRETRAINED.PATH $YOUR_PRETRAINED_WEIGHTS_PATH \
                BATCHSIZE 1

This will render predictions from the network and save them to an visualization_outputs folder.

License

PowerBEV is released under the MIT license. Please see the LICENSE file for more information.

Credits

This is the official PyTorch implementation of the paper:

PowerBEV: A Powerful yet Lightweight Framework for Instance Prediction in Bird's-Eye View
Peizheng Li, Shuxiao Ding,Xieyuanli Chen,Niklas Hanselmann,Marius Cordts,Jürgen Gall