YOLOv2 : Object Detection

YOLOv2 is a state of the art Object Detection Neural Network. This repository is an implementation of YOLOv2 from scratch in tensorflow (eager).

There are two different versions of this Network:

1. Trained on Pascal VOC 2012
2. Trained on Multiview 3D Hand Pose Dataset

The weights for these can be downloaded from Model checkpoint for Object Detection VOC and Model checkpoint for Hand Detection.

Result

Requirements

1. Python 3.5+
2. Tensorflow version 1.7.0 and its requirements.
3. NumPy 1.9.0.
4. OpenCV 3.4.0 The opencv-python pip should work
5. CUDA version 9 (Strongly Recommended)
6. cuDNN version 7.0 (Recommended)

Datasets

The datasets used to train the network are listed below:

Name of Dataset	No. of Images	No. of Classes
Pascal VoC 2013	17125	20
Multiview 3D Hand Pose Dataset	Around 700K	1

Network Architecture

	Layer	Filters	Size	Input Layer	Input Dimension	Output Dimension
1	Conv 2D	32	3 x 3 / 1	Input	416 x 416 x 3	416 x 416 x 32
1	Max Pool		2 x 2 / 2	1 Conv 2D	416 x 416 x 32	208 x 208 x 32
2	Conv 2D	64	3 x 3 / 1	1 Max Pool	208 x 208 x 32	208 x 208 x 64
2	Max Pool		2 x 2 / 2	2 Conv 2D	208 x 208 x 64	104 x 104 x 64
3	Conv 2D	128	3 x 3 / 1	2 Max Pool	104 x 104 x 64	104 x 104 x 128
4	Conv 2D	64	3 x 3 / 1	3 Conv 2D	104 x 104 x 128	104 x 104 x 64
5	Conv 2D	128	3 x 3 / 1	4 Conv 2D	104 x 104 x 64	104 x 104 x 128
5	Max Pool		2 x 2 / 2	5 Conv 2D	104 x 104 x 128	52 x 52 x 128
6	Conv 2D	256	3 x 3 / 1	5 Max Pool	52 x 52 x 128	52 x 52 x 256
7	Conv 2D	128	3 x 3 / 1	6 Conv 2D	52 x 52 x 256	52 x 52 x 128
8	Conv 2D	256	3 x 3 / 1	7 Conv 2D	52 x 52 x 128	52 x 52 x 256
8	Max Pool		2 x 2 / 2	8 Conv 2D	52 x 52 x 256	26 x 26 x 256
9	Conv 2D	512	3 x 3 / 1	8 Max Pool	26 x 26 x 256	26 x 26 x 512
10	Conv 2D	256	3 x 3 / 1	9 Conv 2D	26 x 26 x 512	26 x 26 x 256
11	Conv 2D	512	3 x 3 / 1	10 Conv 2D	26 x 26 x 256	26 x 26 x 512
12	Conv 2D	256	3 x 3 / 1	11 Conv 2D	26 x 26 x 512	26 x 26 x 256
13	Conv 2D	512	3 x 3 / 1	12 Conv 2D	26 x 26 x 256	26 x 26 x 512
13	Max Pool		2 x 2 / 2	13 Conv 2D	26 x 26 x 512	13 x 13 x 512
14	Conv 2D	1024	3 x 3 / 1	13 Max Pool	13 x 13 x 512	13 x 13 x 1024
15	Conv 2D	512	3 x 3 / 1	14 Conv 2D	13 x 13 x 1024	13 x 13 x 512
16	Conv 2D	1024	3 x 3 / 1	15 Conv 2D	13 x 13 x 512	13 x 13 x 1024
17	Conv 2D	512	3 x 3 / 1	16 Conv 2D	13 x 13 x 1024	13 x 13 x 512
18	Conv 2D	1024	3 x 3 / 1	17 Conv 2D	13 x 13 x 512	13 x 13 x 1024
19	Conv 2D	1024	3 x 3 / 1	18 Conv 2D	13 x 13 x 1024	13 x 13 x 1024
20	Conv 2D	1024	3 x 3 / 1	19 Conv 2D	13 x 13 x 1024	13 x 13 x 1024
21	Conv 2D	64	3 x 3 / 1	13 Conv 2D	26 x 26 x 512	26 x 26 x 64
22	Conv 2D	1024	3 x 3 / 1	20 + 21	13 x 13 x (1024 + 4*64)	13 x 13 x 1024
23	Conv 2D	125 = 5 x (1+4+20)	3 x 3 / 1	22 Conv 2D	13 x 13 x 1024	13 x 13 x 125

First time setup

git clone https://github.com/pmkalshetti/object_detection.git
cd object_detection
pip install -r requirements.txt

Model - Generating data, Training and Prediction

• Make necessary changes in constants.py to reflect your dataset (both train and prediction) then run the following commands to make TFRecords, that will be loaded into the train.py.

    cd src
    python3 write_data_to_TFRecords.py

• Train the model using the following command. This might take a long time.

    python3 train.py

• Do the prediction. If not done yet, make changes to reflect your input to do the prediction. Make changes in constants.py to reflect the output directory for network predictions. Then run the following to do the inference.

    python3 predict.py

This will currently work for the dummy data provided, if you do not make any changes in constants.py.

Results

The mAP achieved with this implementation of YOLO is 0.82.

Here are some visual results obtained by the network.

References

• Original Paper: YOLO9000: Better, Faster, Stronger
• Original paper's authors webpage for weights and other info: Darknet
• A YOLOv2 implementation in Keras: https://github.com/allanzelener/YAD2K
• Dataset: Pascal VOC 2012
• Dataset: Multiview 3D Hand Pose Dataset