Fix output image results in Semantic Segmentation task from run.py #889
Conversation
|
This PR needs Approvals as follows.
Please choose reviewers and requet reviews! Click to see how to approve each reviewsYou can approve this PR by triggered comments as follows.
See all trigger commentsPlease replace [Target] to review target
|
| img = img.convert("L") | ||
| img = np.zeros(label.shape[:2], dtype=np.uint8) | ||
| img[label == i] = 255 | ||
| img = PIL.Image.fromarray(img, mode="L").convert('1') |
There was a problem hiding this comment.
@oatawa1
Sorry, I don't understand why this diff solve the problem
Can you teach me?
There was a problem hiding this comment.
The new commit assigns each pixel color by the predicted class (the highest probability from softmax), Thus, it can be seen as assign 255 to the pixel that belongs to that class.
$ \sum_{\for i=class_no}{max(p_i) * 255} $
While each pixel in the old commit is a combination of each class based on softmax probability.
$ \sum_{\for i=class_no}{p_i * 255} & sum(p_i) = 1.0 $
|
/ready |
|
⏳Merge job is queued... |
What this patch does to fix the issue.
Resolve the broken output image in Semantic Segmentation task from
run.pyLink to any relevant issues or pull requests.
#886 #830