Faster R-CNN on BDD100K using PyTorch

Images along with bounding boxes and scores present in BDD100K dataset for object detection

Installation

Its advisible to create an conda environment

conda create --name env
conda activate env

Install the latest version from source code:

$ git clone https://github.com/anupamkliv/bdd100k.git
$ pip install -r requirements.txt

Using Docker

Create a docker image

 docker build . -f Dockerfile.py -t #image_name .

Provide #image_name as required

Run the docker image

docker run #image_name

In line 17 of the docker file use the python file name for which the docker image needs to be made.

Documentation

This repo consist of 3 three things

Dataset overview
Training
Inference

Training and inference are done using the same train.py file

1. Dataset overview

python dataset.py \
 --all_stats 1 \
 --train_stats_attribute 0 \
 --imbalance 0 \
 --unclear_image 0 \
 --traffic_light 0 \
 --area_stats 0 \
 --visualize_dataset 0 \
....

2. Training and inference

python train.py \
 --lr 0.001 \
 --batch_size 8 \
 --epochs 2 \
 --model fasterrcnn_resnet50_fpn \
 --train 1 \
 --iou_thres 0.5 \
 --prob_thres 0.02 \
....

Arguments during Dataset Overview and Training& Inference

Dataset Overview

Argument	Description	Default
all_stats	if 1 it will give all the statistics	1
train_attribute	get no of images of each attribute from train set	0
valid_attribute	get no of images of each attribute from valid set	0
imbalance	gives class and sample imbalance statistics	0
unclear_image	gives number of blur images	0
traffic_light	gives where trafficlightcolor is None	0
area_stats	determines the mean and std of areas in the bounding box (BB)	0
visualize_dataset	original and bounding box visualization	0
number_visualize	number of images to be visualized	3
train_json	label file for training dataset	dataset/bdd100k_labels_release/bdd100k/labels/bdd100k_labels_images_train.json
label_json	label file for validation dataset	dataset/bdd100k_labels_release/bdd100k/labels/bdd100k_labels_images_val.json
train_images	path for train images	dataset/bdd100k_images_100k/bdd100k/images/100k/train
label_images	path for validation images	dataset/bdd100k_images_100k/bdd100k/images/100k/val

Training & Inference

Argument	Description	Default
lr	specifies learning rate used during training	0.001
batch_size	specifies dataset batch size	8
epochs	specifies epochs during training	2
model	specifies network architecture	fasterrcnn
img_dir	specifies image directory	dataset/bdd100k_images_100k/bdd100k/
label_dir	specifies label directory	dataset/bdd100k_labels_release/bdd100k/
train	specifies whether to train or only inference	1
iou_thres	specifies the IoU threshold taken for mAP	0.5
prob_thres	specifies score threshold to visualize BB	0.02
train_images	specifies the number of images to be taken during training	50
val_images	specifies the number of images to be taken during inference	3

Architecture

Files architecture of this repo. These contents may be helpful for users to understand our repo.

|── bdd100k
│   ├── configs
│   │   ├── test_eval.json 
│   │   └── test_train.json
│   ├── dataset
│   │   ├── bdd100k_images_100k
│   │   └── bdd100k_labels_release
│   ├── dataset_info
│   │   ├── dataset_info.py
│   │   ├── plot.py
│   │   └── utils.py
│   ├── test
│   │   └── unittests.py
│   ├── utils
│   │   ├── dataset.py
│   │   ├── evaluation.py
│   │   ├── model.py
│   │   ├── trainer.py
│   │   ├── train_utils.py
│   │   ├── transformations.py
│   │   └── visualize.py
│   ├── dataset.py
│   ├── train.py
│   ├── Dataset_information.ipnyb
│   ├── Dockerfile
│   └── requirements.txt

Download the dataset from: Link and prepare the dataset in the following structure.

|── dataset
│   ├── bdd100k_images_100k
│   │   └── bdd100k
│   │       └── images
│   │           └── 100k
│   │               ├── train
│   │               ├── val
│   │               └── test
│   └── bdd100k_labels_release
│       └── bdd100k
│           └── labels
│               ├── bdd100k_labels_images_train.json
│               └── bdd100k_labels_images_val.json

Running tests

An example of test case is available in the test directory. To run the test, run the following command from the root directory:

python -m test.unittests

Datasets Supported

Dataset	Training samples	Validation samples	Classes
BDD100k	70,000	10,000	10

Data Partition

We provide prospect of the dataset. Here we show snapshot for some of the statistics.

Frequency of classes present in training and validation set

Number of images present in training and validation w.r.t classes

Models Supported

This repo has support for the following Deep Learning model, which are loaded from torchvision.models:

FasterRCNN Torchvision

Dominance of Car, Traffic sign and Traffic Light

Random 3 images taken from Training set

Random 3 images taken from Validation set

Incorrect ground truth

Validation set

References

[1] Ren, Shaoqing, et al. "Faster r-cnn: Towards real-time object detection with region proposal networks." Advances in neural information processing systems 28 (2015). Link

Contact

For technical issues, please contact our development team through Github issues or email:

Contributor

Anupam Borthakur
Centre of Excellence in Artificial Intelligence,
Indian Institute of Technology Kharagpur
email: anupamborthakur@kgpian.iitkgp.ac.in
Github username: anupamkliv