Example doesnt work

[mrinath123]

📚 The doc issue

In this example for fasterrcnn

vision/torchvision/models/detection/faster_rcnn.py

Line 354 in e65a857

Example::

in the # For training part
the output is giving an error
"ValueError: All bounding boxes should have positive height and width. Found invalid box [0.45278316736221313, 0.678650438785553, 1.0717594623565674, 0.416100412607193] for target at index 0."

Suggest a potential alternative/fix

Maybe change the way bbox are defined, here it is being initialized randomly

[datumbox]

@mrinath123 Thanks, you are right. Executing the following:

import torch
import torchvision

model = torchvision.models.detection.fasterrcnn_resnet50_fpn(pretrained=True)
images, boxes = torch.rand(4, 3, 600, 1200), torch.rand(4, 11, 4)
labels = torch.randint(1, 91, (4, 11))
images = list(image for image in images)
targets = []
for i in range(len(images)):
    d = {}
    d['boxes'] = boxes[i]
    d['labels'] = labels[i]
    targets.append(d)
output = model(images, targets)
 # For inference
model.eval()
x = [torch.rand(3, 300, 400), torch.rand(3, 500, 400)]
predictions = model(x)

Leads to the above exception because some of the bottom right coordinates of the boxes have larger values than the ones of top left. As you correctly pointed out, to fix this we need to ensure that boxes[:, :, 2:4] > boxes[:, :, 0:2].

We could add something like the following right after the initialization of boxes to fix it but other patches exist:
boxes[:, :, 2:4] = boxes[:, :, 0:2] + boxes[:, :, 2:4]

Let me know if you would like to send a PR to fix the problem.

[mrinath123]

I will love to send you a PR, would be my 1st contribution to this awesome framework!

[datumbox]

Sounds great! Feel free to add me as a reviewer once you send the PR. Looking forward to your contribution. :)

[mrinath123]

Hi! I tried to make a PR #5165 it's my first time contributing, if I make some mistake do tell me.
Thanks