\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1377382%2Fca7d1c6daf41399ece408387f7d630fc%2FScreenshot%20from%202021-01-12%2014-11-30.png?generation=1610441110509854&alt=media\")
In cases where a particular pathy occurs only once but has multiple annotations by different radiologits, an intercept can be taken easily. (see the example below)
\n
But it starts getting complicated when there are multiple instances of the same pathology in the same image, each having different annotations by radiologists. (see example below)
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1377382%2F65830e06f247d1dcdee936508d52c1c8%2FScreenshot%20from%202021-01-12%2014-03-32.png?generation=1610441189145939&alt=media\")
And then there's this monstrosity:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F1377382%2F7b9b5edbfa7b9b3b571ec96b8512a269%2FScreenshot%20from%202021-01-12%2001-16-11.png?generation=1610441371938145&alt=media\")
What's the best way to handle multiple overlaps with multiple occurrences in the same image?
", "messages": [ { "id": 1149925, "postDate": "2021-01-12T08:50:52.060Z", "content": "In cases where a particular pathy occurs only once but has multiple annotations by different radiologits, an intercept can be taken easily. (see the example below)
\n
But it starts getting complicated when there are multiple instances of the same pathology in the same image, each having different annotations by radiologists. (see example below)
\n
And then there's this monstrosity:
\n
What's the best way to handle multiple overlaps with multiple occurrences in the same image?
", "rawMarkdown": "In cases where a particular pathy occurs only once but has multiple annotations by different radiologits, an intercept can be taken easily. (see the example below)\n\n\nBut it starts getting complicated when there are multiple instances of the same pathology in the same image, each having different annotations by radiologists. (see example below)\n\n\nAnd then there's this monstrosity:\n\n\nWhat's the best way to handle multiple overlaps with multiple occurrences in the same image? \n", "votes": 7 }, { "id": 1153156, "postDate": "2021-01-14T16:44:19.863Z", "content": "This notebook explores some of the methods for combining bounding boxes such as Non-maximum Suppression (NMS), Soft-NMS, Non-maximum Weighted (NMW) and Weighted Bounding Box Fusion.
\nThere's also the question on whether for some class_idvalues you should only ever allow 1 bounding box (that may be relevant both for combining bounding boxes before training, as well as as a possible post-processing for predictions).
Finally, I've wondered to what extent just randomly selecting the annotations by a different radiologist in each epoch would be a sensible augmentation strategy. Here is a prototype dataloader that does that when reading the data from a shelve file I created in another notebook the idea with using shelve is to speed things up vs. reading the .dicom files again, and again for each batch.
This notebook explores some of the methods for combining bounding boxes such as Non-maximum Suppression (NMS), Soft-NMS, Non-maximum Weighted (NMW) and Weighted Bounding Box Fusion.
\nThere's also the question on whether for some class_idvalues you should only ever allow 1 bounding box (that may be relevant both for combining bounding boxes before training, as well as as a possible post-processing for predictions).
Finally, I've wondered to what extent just randomly selecting the annotations by a different radiologist in each epoch would be a sensible augmentation strategy. Here is a prototype dataloader that does that when reading the data from a shelve file I created in another notebook the idea with using shelve is to speed things up vs. reading the .dicom files again, and again for each batch.