Why do all the "score" outputs come out zero?

[machengcheng2016]

Hello! Came here from your published paper! Impressive work of this project I have to say! But here are two questions to ask... First, I test the provided pre-trained model "ern_UAV.caffemodel" on the UAV dataset, but all the "score" results come out as black images, which means all the segmentation labels are zero. I wonder why is this happening. Maybe the preprocessing code in "UAV_infer.py" is not proper? Second, specifically, I'm using your "UAV_infer.py" to run the test, and is the "transform the input image" part necessary? As a result, could you please provide a little piece of test data and an exact Python script to let me finish the test? My Gmail address is: machengcheng2016@gmail.com Appreciate it!

[liushuo2018]

I'm sorry for the failure of your experiments. First, I have re-tested the code and found it could work properly. Note that the output is "label map" ([0,1,2,3,4,5] for UAV images and VH images ), I guess you may havn't transformed the label map to color map? This is an easy operation. Second, If you use our trained model weights and UAV_infer.py / VH_infer.py, you should first split the orignial images into 256*256 patches (with or without overlap). As the UAV images have not been publicly available, I recommend you test the model with VH data. The difference between the UAV_infer and VH_infer lies: UAV_infer takes images (RGB, 3 channels) as input, VH_infer takes imges + nDSM (CIR+nDSM, 4 channels) as input.

[machengcheng2016]

It's so kind of you to re-test the code! After transforming the label map to color map, I did get colorful label map! The problem is solved! As for the training process of UAV images, I still have two problem: 1) In the beginning of training, I convert image, label and edge data(all these data are prepared as ".jpg" images) to LMDB format data to feed the network. According to the paper, labels should be {0,1,2,3,4,5} and edges should be {0,1}. So is it proper to prepare labels and edges as ".jpg" gray images? Since the range of an image is [0,255], the labels may turn into {0,51,102,153,204,255} and edges turn into {0,255}...

2) The first problem really confuses me, and I finally decide to convert image, label and edge data to HDF5 data format. The HDF5 format can store float data without the restrict of [0,255]. So this time labels and edges are exactly prepared as {0,1,2,3,4,5} and {0,1}. But for images, I still scale images to [0,255] integers. Should I scale images to [0,1] float number? How do you judge my process? Thank you for your help!

[liushuo2018]

Use .png rather than .jpg may fix your problem.

[machengcheng2016]

Thanks!