Add widerface dataset

[jgbradley1]

This PR addresses issue #1627 and adds a dataset that is more representative (and harder) of generic face detection.

Main changes in the PR include

• adds WIDERFace dataset as an option
• improves the method to check and download files from google drive when download quota has been exceeded. See the Google drive API docs for more details, A quota limit exceeded will return a 403 response. Checking for the correct response code before performing a string search saves a significant amount of time for the most general use case (i.e. when the quota limit has not been exceeded).

cc: @pmeier @vfdev-5

[codecov]

Codecov Report

Merging #2883 (c11858f) into master (36daee3) will decrease coverage by 0.84%.
The diff coverage is 51.31%.

@@            Coverage Diff             @@
##           master    #2883      +/-   ##
==========================================
- Coverage   73.41%   72.57%   -0.85%     
==========================================
  Files          99      100       +1     
  Lines        8801     8947     +146     
  Branches     1389     1419      +30     
==========================================
+ Hits         6461     6493      +32     
- Misses       1915     2033     +118     
+ Partials      425      421       -4     

Impacted Files	Coverage Δ
torchvision/datasets/folder.py	85.54% <ø> (ø)
torchvision/models/detection/anchor_utils.py	89.61% <0.00%> (-2.50%)	⬇️
torchvision/models/quantization/inception.py	31.45% <0.00%> (ø)
torchvision/datasets/widerface.py	15.70% <15.70%> (ø)
torchvision/models/inception.py	85.03% <33.33%> (+0.39%)	⬆️
torchvision/models/detection/retinanet.py	75.87% <50.00%> (-0.21%)	⬇️
torchvision/datasets/__init__.py	100.00% <100.00%> (ø)
torchvision/models/densenet.py	83.58% <100.00%> (+0.76%)	⬆️
torchvision/models/detection/_utils.py	85.43% <100.00%> (+0.39%)	⬆️
torchvision/models/detection/backbone_utils.py	95.55% <100.00%> (+1.26%)	⬆️
... and 11 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 36daee3...1f0223c. Read the comment docs.

[jgbradley1]

I think it would be preferable instead to always return the "raw" mode as a dict. I know we have target_type in CelebA already, but I would prefer if the output type of the datasets was always the same.

@fmassa I need some clarification. Are you suggesting to remove the target_type argument entirely and always return a dictionary? For example, the raw annotation for one image with one face would be:

{
  "raw": tensor([[983, 296, 7, 9, 2, 0, 0, 1, 1, 0]]), # torch.Size([1, 10])
  "bbox": tensor([[983, 296, 7, 9]]),                  # torch.Size([1, 4])
  "attributes": tensor([[2, 0, 0, 1, 1, 0]])           # torch.Size([1, 6])
}

Should the raw key-value pair be left as part of the dict? The tensors will change shape based on the number of faces in an image but hopefully you get the idea from the example above.

If target_type is removed, this may change how target_transform is handled. I used CelebA as a template for applying target_transform. Is there a better Dataset class I should reference as the standard?

[jgbradley1]

@fmassa and @pmeier I'm just picking this PR back up. It got lost during the holidays I think.

I've removed the dataset citation for now and removed the target_type argument as requested earlier. Also updated the output type to always be a dict.

Do you have any other suggestions or is this about ready to merge?

[pmeier]

If I understand your implementation correctly, the target is always returned as None for the test split. If that is the case, I would like to hear @fmassa s thoughts about this, since it would create precedence. As of now, all our datasets always return a valid target.

Other than that two minor nitpicks.

[jgbradley1]

If I understand your implementation correctly, the target is always returned as None for the test split. If that is the case, I would like to hear @fmassa s thoughts about this, since it would create precedence. As of now, all our datasets always return a valid target.

Ground truth annotations for the WIDERFace test split are not released publicly (as noted in the description section of the main website). In this case, the only options we can provide here is either:

1. drop support for the test split or
2. support reading the test split (to improve ease of evaluation) and return an empty target

A comment has been added to getitem to explain the empty target to users.

Actual evaluation on the test split requires a submission to the dataset author. I prefer option 2 as it would encourage users to go for SOTA on the dataset.

[fmassa]

Hi @jgbradley1

Sorry for the delay.

I think I would be ok letting the dataset return a target=None for the test split. This way, the return types are always consistent.

If we were to return either image or image, target, then I would be a bit more reticent to do it, but returning image, None I think it's fine.

[jgbradley1]

Looks like there is a problem with a couple of the conda CI builds (unrelated to my changes as far I can tell). I don't think it's anything I can resolve from my end.

[pmeier]

Thanks a lot @jgbradley1!