Deeplab does not produce correct segmentation mask.

[ajinkya933]

---

System information

• What is the top-level directory of the model you are using: models->research ->Deeplab
• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): Yes
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 16.04
• TensorFlow installed from (source or binary): source
• TensorFlow version (use command below):1.13
• Bazel version (if compiling from source): bazel-0.24.0-installer-darwin-x86_64

Describe the problem

I am training my own model to segment squares in images using Deeplab. For that I have annotated the training data, Heres an example of training image(on left) and segmentation mask image(on right).

I am converting this training data to record file using: Deeplab-> build_voc2012_data.py

I have further made following changes to the scripts described as below, These changes are such that I can train my own dataset.

deeplab -> datasets -> datagenerator.py

TESTSET_INFORMATION = DatasetDescriptor( splits_to_sizes={ 'train': 800, # num of samples in images/training 'val': 200, # num of samples in images/validation }, num_classes=2, ignore_label=255, # When setting ignore_label = 0 predictions come back as a red or green image; and when setting ignore_label = 255 predictions come back as a black image ) _DATASETS_INFORMATION = { 'cityscapes': _CITYSCAPES_INFORMATION, 'pascal_voc_seg': _PASCAL_VOC_SEG_INFORMATION, 'ade20k': _ADE20K_INFORMATION, 'testset': _TESTSET_INFORMATION, }

utils > train_utils.py

exclude_list = ['global_step', 'logits']

not_ignore_mask = tf.to_float(tf.equal(scaled_labels, 0)) 1 + tf.to_float(tf .equal(scaled_labels, 1)) 2 + tf.to_float(tf.equal(scaled_labels, ignore_label )) * 0

Please note batch size=12. And training number of steps = 1000. I get following predictions

I am not sure what is it I am doing wrong ? To further debug the problem I observed that during training the loss dosen't change it keeps on oscillating.

example loss is:

Total loss is :[0.0697127208] INFO:tensorflow:global_step/sec: 0.715997 Total loss is :[0.0697126836] INFO:tensorflow:global_step/sec: 0.835336 Total loss is :[0.0697126538] INFO:tensorflow:global_step/sec: 0.83706 Total loss is :[0.0697126389] INFO:tensorflow:global_step/sec: 0.833028 Total loss is :[0.069712624] INFO:tensorflow:global_step/sec: 0.836747 Total loss is :[0.069712624] INFO:tensorflow:global_step/sec: 0.831312 Total loss is :[0.069712624] INFO:tensorflow:global_step/sec: 0.832811 Total loss is :[0.069712624] INFO:tensorflow:global_step/sec: 0.83014 Total loss is :[0.069712624] INFO:tensorflow:global_step/sec: 0.830697

What should I do so that the loss is reduced and the model gives a proper square as a segmentation output ?

@bleedingfight @aquariusjay @Citygity could you shed some light and help me bring one step closer to the solution ?

[vijay1131]

@ajinkya933 How did you solved your issue?

[hakS07]

@ajinkya933 did you solved the problem??

[ajinkya933]

My input images which I used for training were not used properly, Make sure that your object has pixel value of RGB(1,1,1) and your background has pixel value(0,0,0)

[hakS07]

@ajinkya933 ok, thank you