DeepLab: FAQ


Q: The label read by IMAGE_SEG_DATA is not the same as expected. Do you pre-process the annotations?

A: You need to save the png annotations without the colormap; otherwise OpenCV will read the wrong values. See the script at matlab/my_script/SavePngAsRawPng.m for reference.


Q: I am finetuning the VGG 16-layers.caffemodel with your prototxt, I have a fatal error when loading fc6 parameters.

A: We did not use the original VGG-16; we changed the kernel size from 7×7 to 4×4 or 3×3 and employed the atrous algorithm. Please download the provided init models at here.


Q: When compiling the codes, I encounter this error message: “libmatio.so: undefined reference”.

A: Please install matio.


Q: A folder called /SegmentationClassAug/ does not exist in the original PASCAL VOC2012 Devkit.

A: You need to download the extra annotations provided by Bharath Hariharan et al. at here. Please consider citing their paper if using the dataset.


Q: When training the model, I have encountered this error message: “[softmax_loss_layer.cpp:85] Unexpected label 255″.

A: Please add “ignore_label: 255″ to both softmax_loss_layer and seg_accuracy_layer. See example here.


Q: When evaluating the DeepLab outputs (without CRF), I got displaced results (i.e., results not aligned with image). Is there anything wrong?

A: This happens to several users, and it is resolved after compiling without cuDNN enabled. Maybe you could try that too.


Q: When evaluating the DeepLab outputs (without CRF), I got all-background results (i.e., all black results). Is there anything wrong?

A: Please double check if the name of your fc8 is fc8_voc12 in the generated test_val.prototxt or test_test.prototxt (after running run_pascal.sh). The name should be matched for initialization.