Yolo v3 Tiny (TensorFlow 2.x) for noisy mnist detection

It turns out the trained NN from a clean background basically does not work on noisy testing set (even with only a little noise, e.g., mean = variance = 0.1). In this project, a Yolo-v3-tiny model was re-trained (from the mnist dataset with Gaussian additive noise) to detect random digits appearing in a noisy background.

Installation

First, clode or download this GitHub repository. For Windows, put wget.exe under the project folder.

Install requirements and download pretrained weights, CMD open console:

pip install -r ./requirements.txt

# yolov3 (not needed here)
wget -P model_data https://pjreddie.com/media/files/yolov3.weights

# yolov3-tiny
wget -P model_data https://pjreddie.com/media/files/yolov3-tiny.weights

Quick demo

Start with using pretrained weights to test predictions on both image and video:

python show_detect_performance.py

Prepare a noisy mnist dataset

mnist folder contains mnist images, create training data:

python mnist/make_data.py
python mnist/add_noise2images.py

You can change training hyperparameters here ./yolov3/configs.py.

Custom training of Yolo v3 object detection

I highly recommend to train the following step on Google Colab (see the last section below).

I could use my laptop to fry eggs after 100 epochs training (~3 hours) with a GPU Quadro M1200 (computing power of 5).

python train2.py

Test detection with detect_mnist.py script:

python detect_mnist2.py

Results:

Google Colab Custom Yolo v3 training

To learn more about Google Colab Free gpu (Tesla T4 with the computing power 7!) training, visit this tutorial and demo notebook (for using Colab) in this project.