Improved-JNetV3 SSD in Pytorch

This repo is forked from pytorch-ssd.

• We use MobileNetV2 as basenet
• Change image input size form 300x300 to 768x768, it is propitious to the details of table
• We use improved JNetV3 segmentation network, treated as a Auxiliary Supervisory Signal.It can accelerate the training process(get same loss : 1 night->20 min) and get a better accuracy
• Change anchor for long width with short height table ceil

Improved-JNetV3 SSD architecture

original MobileNetV2 SSD

Improved-JNetV3 SSD

Improved-JNetV3 SSD Heatmap

Dependencies

1. Python 3.6+
2. OpenCV
3. Pytorch 0.4+
4. Caffe2
5. Pandas
6. Boto3 if you want to train models on the Google OpenImages Dataset.

Train

train_ssd.py #train file

Test

eval_ssd.py #test file