使用 Tensorflow deeplab 可视化我自己的一组图像

Question

I am trying to implement Deeplab example from Tensorflow. I followed the guideline and manage to train it with Cityscapes dataset with accuracy of 0.77. I am able to use the vis code to create segmented images with the images of the Cityscape dataset.

Now I want to use my set of images to be visualized, I tried replacing the files from the dataset by putting them in the folder /models/research/deeplab/datasets/cityscapes/leftImg8bit/val and rerun sh convert_cityscapes.sh but it finishes creating tfrecord when I run

python deeplab/vis.py 
 --logtostderr \
 --vis_split="val" \
 --model_variant="xception_65" \
 --atrous_rates=6 \
 --atrous_rates=12 \
 --atrous_rates=18 \
 --output_stride=16 \
 --decoder_output_stride=4 \
 --vis_crop_size=1025 \
 --vis_crop_size=2049 \
 --dataset="cityscapes" \
 --colormap_type="cityscapes" \
 --checkpoint_dir=${PATH_TO_CHECKPOINT} \
 --vis_logdir=${PATH_TO_VIS_DIR} \
 --dataset_dir=${PATH_TO_DATASET}

It doesnt create anything.

I dont need to run the training again, I just want to use my pretrained model to predict the image segmentations from my own images, but I dont know how to progress.

Answer 1

Did you make sure the tfrecords you created are of pattern: val-* ? You can also create your own DatasetDescriptor so that you won't mix up with your true val set.