Yet-another-efficientdet-pytorch: [Customer Small Object Dataset ] If you have not modified your anchor, do give it a try

[Cli98]

Hi guys,

I explored how anchor works in past three weeks. And actually it is much more important than I expected.

Just give you guys some examples.

If you see green boxes, then it is a match. Else you get A BIG ISSUE. For small object detection dataset or aerial image dataset, I guess that you all need to recompute anchor parameters.

Give me some comments on my work if you can, thank you! Free feel to email me ([email protected]) for personal/professional discussion.

[FancyFai]

Hi guys,

I explored how anchor works in past three weeks. And actually it is much more important than I expected.

Just give you guys some examples.

If you see green boxes, then it is a match. Else you get A BIG ISSUE. For small object detection dataset or aerial image dataset, I guess that you all need to recompute anchor parameters.

Give me some comments on my work if you can, thank you! Free feel to email me ([email protected]) for personal/professional discussion.

Hello, I would like to ask whether you have modified anchors_ratios and anchor_scale. If I get the height and width of anchor through kmeans（like [[358 ,144] [405, 414] [550, 637]]）, how to convert them to the corresponding anchors_ratios and anchor_scale? Can you give me some Suggestions. Looking forward to your reply

[Cli98]

Hi guys,
I explored how anchor works in past three weeks. And actually it is much more important than I expected.
Just give you guys some examples.


If you see green boxes, then it is a match. Else you get A BIG ISSUE. For small object detection dataset or aerial image dataset, I guess that you all need to recompute anchor parameters.
Give me some comments on my work if you can, thank you! Free feel to email me ([email protected]) for personal/professional discussion.

Hello, I would like to ask whether you have modified anchors_ratios and anchor_scale. If I get the height and width of anchor through kmeans（like [[358 ,144] [405, 414] [550, 637]]）, how to convert them to the corresponding anchors_ratios and anchor_scale? Can you give me some Suggestions. Looking forward to your reply

There are already lots of discussion on this topic in current repo. In short, you need a two-factor decomposition. To summarize, 1). Use kmean to find suitable anchors. 2). Compute scale and ratio from kmean result. Make sure scale*ratio -> kmean result.

I will post my code for anchor computations after I pass necessary tests. Stay tuned.

[jinfagang]

@Cli98 Can u post your kmeans and scale/ratio calculation code?

[Cli98]

@Cli98 Can u post your kmeans and scale/ratio calculation code?

I'm testing my code now. If I am able to see Improvements in AP then I will modify it and post here.

[jinfagang]

@Cli98 Why closing issue?

[Cli98]

@Cli98 Why closing issue?

@jinfagang Just get notified. As there are already lots of issues in this repo. I do not want to take other's slot. I will PR when I'm ready to share the code.

[Cli98]

@Cli98 Why closing issue?

@jinfagang Just get notified. As there are already lots of issues in this repo. I do not want to take other's slot. I will PR when I'm ready to share the code.

@jinfagang Give it a try if you need. https://github.com/Cli98/anchor_computation_tool
Good luck!