code for visualization in the object detection tutorial

[isalirezag]

At the end of the object detection tutorial it visualizes the masks.
can you please provide the code for that task? or guide how to do it?

[fmassa]

Hi,

The code I used for the masks uses OpenCV (which torchvision doesn't use as a dependency), and that's one of the reasons why I didn't share that code yet.

Here are the scripts: https://gist.github.com/fmassa/d802c684512bcc99f54b714f10c96872

I used something like the following to visualize

model = torchvision.models.detection.keypointrcnn_resnet50_fpn(pretrained=True)
model.cuda()
model.eval();

# img is a PIL Image
result, output, top_pred = predict(img, model)
plt.imshow(Image.fromarray(result[:, :, [2, 1, 0]]))

Let me know if you have questions. I think it makes sense to add those functions somewhere in torchvision, but I'm not yet sure where

[mattans]

@fmassa , since you used a pretrained model, didn't you need to normalize your PIL image before?

[fmassa]

Hi @mattans

Normalization for the detection models happen inside the model, see

vision/torchvision/models/detection/transform.py

Line 104 in 9b80465

image = self.normalize(image)

[mattans]

Great, thanks.
Linking to #2397 (comment) as it's related.