4
votes
Accepted
Why is my training loss not changing?
Dying ReLU is a good guess. But the question is why this is not happening in original paper. Answer is input data features. Have you normalize the data? You can debug this issue by trying to find the ...
- 188
3
votes
Accepted
What is Deep supervision?
Partial answer: Quoting a paper
"The advantage of such deep supervision is evident: (1)for small training data and relatively shallower networks, deep supervision functions as a strong “...
- 5,759
2
votes
Is there an official procedure to compute mIoU?
I came across this very question myself while I was trying to replicate the results reported on a paper relevant to semantic segmentation.
It turns out there is not really a universal way of computing ...
- 188
2
votes
How do you train a semantic segmentation model to optimize for IoU rather than accuracy?
You could directly optimize the mean IoU loss by implementing the following loss:
...
- 2,792
2
votes
What is Deep supervision?
The idea of deep supervision is to add, so called, companion objective functions at each hidden layer of a network and then compute the final loss as the output loss plus the sum of the companion ...
- 121
2
votes
Accepted
Why not using segmentation architectures for object detection?
At the end of the day, if an architecture satisfies your goal, you should use it. If you can detect objects using a segmentation model, go ahead.
But that's where practice differs from theory. Don't ...
- 6,106
1
vote
Accepted
How do I ensure final output shape matches input shape for a semantic segmentation task?
The issue comes from the fact that a max pooling operation is applied to downsample at each level of the U-Net. When the your input is not divisible by two the resulting array after max pooling will ...
- 8,197
1
vote
Building a CNN (with Keras for pixelwise classification)
The last layer should match the dimension of your response. Dense layer can only return 1D whereas conv layers can return 2D.,
Based on your question, you want to classify all your pixels (binary?) in ...
- 31
1
vote
Why not using segmentation architectures for object detection?
In addition to the great answer by Valentin there is the aspect of the data labeling.
Assume you want to tackle a new task of a specific object.
Labeling ~10,000 samples with bounding box is doable in ...
- 176
1
vote
Accepted
Can I use one-hot encoded output for segmentation in Pytorch, with focal and dice losses?
For classification, it seems that in the latest version of PyTorch, cross-entropy accepts one_hot encoded labels as well.
For segmentation, PyTorch does not have a Dice loss implementation, hence it ...
- 121
1
vote
Accepted
Remove frame from background
MODNet, based on the paper, is a network that removes the background entirely just as if the person is standing in front of a green screen. And I assume that is why you are left only with the player ...
- 120
1
vote
Tool for annotation of images for semantic segmentation
Diffgram is really great for this! I used it for a construction monitoring project. It's Open Source. From their site:
Semantic Segmentation Tools:
Auto Bordering: Automatically detects edges to ...
- 111
1
vote
How to extract contents by topic from a document?
Here are a list of tools you can look into:
https://tika.apache.org/
https://jsoup.org/
https://poi.apache.org/
This was a neat read detailing the steps. The author was doing something similar to ...
- 111
1
vote
Accepted
Semantic networks: word2vec?
There are a few models that are trained to analyse a sentence and classify each token (or recognise dependencies between words).
Part of speech tagging (POS) models assign to each word its function (...
- 1,064
1
vote
Accepted
For semantic sementation, why am I getting better loss values with binary cross entropy than dice coef?
So the question asks about why different loss function lead to different error scores.
So globally error is there to help us measure the level of discrimination between the output of the model and the ...
- 1,468
1
vote
resnet50 implementation for semantic segmentation
The output from the ResNet model is a vector containing the probability that the image belongs to each of the n classes, in your case to any of the 21 classes. If ...
- 8,197
1
vote
The channel dimension of the inputs should be defined. Found `None`
I do not see the kernel value set in your code. In Python, a variable has to have a value before being used. For example:
...
- 22.3k
1
vote
My semantic segmentation model classifies everything as background
This answer might be a little late, but I believe that what you need is the Focal Tversky Loss (https://arxiv.org/abs/1810.07842).
Neither the vanilla Focal Loss nor the Dice Loss generalize well for ...
- 11
1
vote
My semantic segmentation model classifies everything as background
Are you using some data augmentation with random crops / rotations / zooms ? If you do, you might have some images with only background labels and if so I would suggest you to add a condition to only ...
- 11
1
vote
Accepted
Suitable instance counting CNN for training on polygonal masks
Which CNN should I use for instance counting given that my dataset consists of labeled polygons?
CNNs for instance segmentation. To start with you can try Mask-RCNN. Here are all state of the art ...
- 864
1
vote
High image segmentation metrics after training but poor results in prediction
I agree the metrics between your test set and validation set are quite close, but looking at your code it seems you may have run for the full 100 epochs.
...
- 388
1
vote
High image segmentation metrics after training but poor results in prediction
The coefficients are reported on your 150 training examples? This looks like a textbook example of overfitting: good performance on training data, bad on test data. The U-net model has a large number ...
- 923
1
vote
Accepted
Criteria for saving best model during training neural network?
The loss is mostly a measure that helps the model learn and is not looked at too much when deciding which model to select. A more business oriented measure is often used for this, e.g. accuracy. Since ...
- 8,197
1
vote
Accepted
Semantic segmentation of an image with multiple labels per pixel
Three separate models (one per channel) will easily learn to predict the >channel category, but it will never output r0g1b2 and r2g3b0 classes in 64 class model because it have never seen these ...
- 49
Only top scored, non community-wiki answers of a minimum length are eligible