Waymo Challenge: Object Detection / Tracking in RGB images*
Frank Gabel¹, Jens Settelmeier²

* Work done during Corona lockdown

¹ Heidelberg University

² KTH Royal Institute of Technology, Stockholm

Summary The waymo dataset is the largest and most diverse autonomous driving datasets ever released. It consists of HD images of 5 cameras (front, front-left, front-right, left, right), LiDAR scans and associated 2D/3D bounding boxes. Data have temporal order (similar to videos). We tackle the problem of 2D object detection by deploying DeepSORT with an EfficientDet backbone.

This is a repo describing our approach to the Waymo Challenge. The challenges mandated to build a model detecting vehicles (anything with wheels, basically), cyclists and pedestrians.

As the data have a pretty significant time component, the general approach was to employ a tracking algorithm with an aggressive association metric, allowing to generalize to occlusions, shape changes etc.

We used a powerful tracking algorithm (DeepSORT) to keep track of objects. Our detection backbone was an EfficientDet with SOTA performance in real time, original paper link: https://arxiv.org/abs/1911.09070.

Requirements

pip install -r requirements.txt

* pytorch==1.1.0 or 1.2.0  
* tqdm  
* opencv-python  
* scipy  
* sklearn
* matplotlib  
* pillow  
* tensorboardX 

Dataset and Weights

• Download: Waymo Open Dataset
• Pretrained weights: Google Drive

Project
|--- EfficientDet-DeepSORT-Tracker
|          |--- main.py
|          |--- train
|                 |--- train_unsupervised.py
|                 |--- ...
|
|--- data
       |--- training
             |--- xxxxxxxxxxxx_0000.tfrecord
             |--- xxxxxxxxxxxx_0001.tfrecord
       |--- test
             |--- yyyyyyyyyyyy_0000.tfrecord
             |--- yyyyyyyyyyyy_0001.tfrecord
             |--- ...

Then, call --data_path='../data'

Hardware

This source code was mainly tested on an NVIDIA 2070 / 1080Ti.

More examples

Run

Using EfficientDet backbone
python run_waymo_deepsort_efficientdet.py --gpu $GPU_TO_USE --p_semi 1.0 --data_path='../data'

Semi-supervised
python run_waymo_deepsort_yolov4.py --gpu $GPU_TO_USE --p_semi 0.5 --data_path='../data'

Train

Training this model entails first training your backbone detector and then training DeepSORT.

Training EfficientDet backbone
python train.py --gpu $GPU_TO_USE --data_path='../data'

Training DeepSORT
python deep_sort/deep/train.py --gpu $GPU_TO_USE --data_path='../data'

Benchmark

If you are not interested in training yourself, you can use our weights. Put them into the base directory and you are good to go.

coefficient	pth_download	GPU Mem(MB)	FPS	Extreme FPS (Batchsize 32)	mAP	mAP 0.1:0.9
D7	efficientdet-weights.pth	~10000	5	-	52.2
Yolo v4	yolov4.weights	~9000	7	-	54.4

References

Appreciate the great work from the following repositories:

• lukemelas/EfficientNet-PyTorch
• signatrix/efficientdet
• vacancy/Synchronized-BatchNorm-PyTorch
• DeepSORT