Can not detect large objects problem

[lucasjinreal]

Hi, I found that If I set crop_size to 400, then I can not detect large objects. even I large crop_size to 600 later, still can not detect large objects.

Do u know why?

[lucasjinreal]

I rechanged cropsize to 600. still can not detect relatively large objects, I am traning detect human only:

[lucasjinreal]

This is very ridiculous: (trained 200 epochs on CrowdHuman, nothing can be detected)

[YonghaoHe]

@jinfagang You may downsample images by 2x or even 3x. You have to know that the detection range of the detector is controlled by config_dict['detection_scales'] = ((4, 20), (20, 40), (40, 80), (80, 160), (160, 320)), namely [4, 320]. The crop_size should be larger than 320 here, but not too large. If crop_size is too large, it will greatly reduce the training speed.

[lucasjinreal]

@YonghaoHe what's the unit of 4, 20..? what's the [4, 320] mean? Pixels? Why fixed it to 320? is that possible add (320, 640) in it?

Actually I tried 400 and 512 and 600 in crop_size, all same behavior. Can not detect human the this size...

[lucasjinreal]

@YonghaoHe Seems change detection_scales involved output_indices:

Any guidlines how to modification?

[lucasjinreal]

@YonghaoHe Any help? I want add a scale.

[YonghaoHe]

@jinfagang I think you do not grasp some basic concepts. [4, 320] (in pixels) is object scale that the detector can detect. Of cause you can set it as your need. crop_size only controls the input size and it should be larger than the upper bound of object scale, say 320 here. But too large input size will greatly reduce the training speed, it is easy to understand. You can see this page for more details

[lucasjinreal]

@YonghaoHe Can u give an example how to add a scale? It need changed a lot of settings....

[lucasjinreal]

@YonghaoHe Also, I don't think scale is the issue of large object can not detect problem. the input resolution is resized, I tried resize my image to 512, still can not detect human in such input resolution..

[lucasjinreal]

@YonghaoHe Can u help?

[lucasjinreal]

@YonghaoHe Can u try train a model on Crowed Human? I think LFD will fail on this dataset.....

[YonghaoHe]

@jinfagang Ok, I will try, you just provide the dataset.

[lucasjinreal]

@YonghaoHe OK, I am simply using CrowdedHuman, can u provide a test on that? Since CrowdedHumna also a dense detection task and very challenge, also, it can test LFD performance and evident it's effectness.

[YonghaoHe]

@jinfagang You mean this dataset?

[lucasjinreal]

@YonghaoHe Yes.

[lucasjinreal]

@YonghaoHe any updates?