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

Dependencies

train_ssd.py #train file

eval_ssd.py #test file

Name		Name	Last commit message	Last commit date
Latest commit History 23 Commits
.idea		.idea
JNetV3		JNetV3
eval_results		eval_results
model_log		model_log
models		models
ocr_table_utils		ocr_table_utils
vision		vision
README.md		README.md
convert_to_caffe2_models.py		convert_to_caffe2_models.py
draw_eval_results.py		draw_eval_results.py
eval_ssd.py		eval_ssd.py
extract_tf_weights.py		extract_tf_weights.py
jnet-ssd-lite-Epoch-1080-Loss-0.6063801158558239.pth		jnet-ssd-lite-Epoch-1080-Loss-0.6063801158558239.pth
open_images_downloader.py		open_images_downloader.py
prune_alexnet.py		prune_alexnet.py
run_ssd_example.py		run_ssd_example.py
run_ssd_live_caffe2.py		run_ssd_live_caffe2.py
run_ssd_live_demo.py		run_ssd_live_demo.py
table_ceils_separate.py		table_ceils_separate.py
table_cells_generate.py		table_cells_generate.py
table_cells_predictor.py		table_cells_predictor.py
test.png		test.png
test_read.py		test_read.py
test_rotate.py		test_rotate.py
train_ssd.py		train_ssd.py
translate_tf_mobilenetv1.py		translate_tf_mobilenetv1.py
visual_tf_models.py		visual_tf_models.py