Imagenet bounding box problems?

[gobob]

Anyone had any issues with bounding box data downloaded from Imagenet? 

When I train a fast-rcnn model with just 2 images and their annotations for one class, it trains the model. When I extend the dataset to 14 images/annotations or bigger eg 500, I get an "assert (boxes[:, 2] >= boxes[:, 0]).all()"

Traceback (most recent call last):

  File "./tools/train_net.py", line 81, in <module>

    roidb = get_training_roidb(imdb)

  File "/home/ubuntu/fast-rcnn/tools/../lib/fast_rcnn/train.py", line 111, in get_training_roidb

    imdb.append_flipped_images()

  File "/home/ubuntu/fast-rcnn/tools/../lib/datasets/imdb.py", line 111, in append_flipped_images

    assert (boxes[:, 2] >= boxes[:, 0]).all()

AssertionError

[Yohai Devir]

Many imagenet annotations are invalid.
The bounding box xmax/ymax values are sometime out of the image boundary.
Note that the values of the image width / height may not be reliable neither.
Yohai

[Xueting Li]

I encountered the same problem, how can I solve this problem?

在 2015年9月16日星期三 UTC+8下午7:38:01，Yohai Devir写道：

[gobob]

There is a solution for the assert issue here.

https://github.com/rbgirshick/fast-rcnn/issues/11