Add Pascal VOC Class Segmentation #37

desimone · 2017-01-21T07:15:41Z

I use the train/val/test split defined the text files found in ./VOCdevkit/VOC2012/ImageSets/Segmentation/*.txt
Perhaps it makes sense to rename the class in line with its associated task like ms-coco does with CocoDetection and CocoCaptions. How about PascalSegmentation?
Should any special considerations or changes be made for semantic segmentation datasets?

torchvision/datasets/pascal.py

+
+    def __getitem__(self, index):
+        img = Image.open(self.images[index]).convert('RGB')
+        target = Image.open(self.masks[index]).convert('RGB')


torchvision/datasets/pascal.py

+        if self.transform is not None:
+            print("transform was not none")
+            img = self.transform(img)
+        # todo(bdd) : perhaps transformations should be applied differently to masks? 


torchvision/datasets/pascal.py

+            return
+
+        # downloads file
+        if os.path.isfile(fpath) and \


torchvision/datasets/pascal.py

+        self.masks = []
+        with open(os.path.join(split_f), "r") as lines:
+            for line in lines:
+                image = os.path.join(image_dir, line.rstrip('\n') + ".jpg")


torchvision/datasets/pascal.py

+        splits_dir = os.path.join(voc_root, 'ImageSets/Segmentation')
+        split_f = os.path.join(splits_dir, 'train.txt')
+        if not self.train:
+            split_f = os.path.join(splits_dir, ' trainval.txt')


fmassa · 2017-01-21T12:22:22Z

Thanks for the PR!
Overall this looks good. I have a few inline comments, and some more generic remarks:

I think we should mention that this is the segmentation dataset. Maybe something like VOCSegmentation?
I'd rather let the user select the split himself, instead on only allowing for train and trainval splits.
This is specific to VOC 2012, maybe it would be good to let the user select between different year contests?
I was planning on adding a VOCDetection dataset as well. Some preliminary version can be found in https://github.com/pytorch/examples/pull/21/files#diff-0344e770fabb635e92d12f8b67f3504a . I'll merge it in when this PR gets merged.

I think we will need to figure out a better way of applying random transforms to both inputs and targets, as discussed in #9 .

desimone · 2017-01-21T19:10:21Z

Thanks for the review, @fmassa. I have a few follow up questions while I wait for the other datasets to download...

I've renamed the class to VOCSegmentation.
Can you elaborate on how you'd want the user to select the split themselves? Should the list be built from all available masks and then shuffled? Sorry, I don't fully track.
I could add multiple years without too much trouble. I'm curious, what would the usecase be for using 2007 over say 2012?
What would be the most idiomatic way to handle errors. At certain points I use asserts. Not sure how I feel about that.
Common operations like archive downloading, extracting, hash checking, file io etc seem common amongst all datasets. Maybe we could abstract that functionality?
I'll defer how to uniformly apply transforms to both mask and images per Random transforms for both input and target? #9.

j-min · 2018-04-27T13:16:04Z

Any updates on this?

fmassa · 2018-04-27T13:18:16Z

@j-min I'll be revisiting this PR before I release my implementation of Detectron. I'm currently in the cleanup process,

fmassa · 2018-12-06T14:20:22Z

Replaced by #663, thanks a lot for the original PR!

desimone added 2 commits January 20, 2017 22:33

initial pascal voc class segmentation support

5fa4caf

Add sanity checking

9091e5a