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.
- 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 byckpt2npy.py
.
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/
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/
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