PSPNet_tensorflow

Introduction

This is a repository forked from hellochick/PSPNet-tensorflow. The code has been updated to TensorFlow 2.1 and adapted for our project. Moreover, eager execution is supported.

Install

• Get restore checkpoint from Google Drive and put into checkpoint directory directly under PSPNet-tensorflow.
• Get the .npy format checkpoint converted from the above .ckpt files from Google Drive and put directly under the project directory. Note: .npy file is created by ckpt2npy.py.

Inference

Eager mode

To get result on single images, use the following command:

python inference_eager.py --img-path=./input/test1.png  

Options:

--checkpoints: path to checkpoint file in .npy format, default ./checkpoint.npy
--flipped-eval
--save-dir: directory to save result, default ./output/

Graph mode

Inference on single image

python inference_graph.py --img-path=./input/test1.npg  

Options:

--checkpoints: path to checkpoint directory in .ckpt format, default ./checkpoint
--flipped-eval
--save-dir: directory to save result, default ./output/

Inference on datasets

To do semantic labeling on the whole dataset MS-COCO 2014 or MegaDepth, use the following command:

python inference_graph_dataset.py --data_path=DATA_PATH --dataset=coco  

where data_path is the same DATA_PATH set in SuperPoint, dataset can be chosen between coco and megadepth. Options:

--flipped-eval