tyomj / product_detection Goto Github PK

hi thanks your research,first I want apologize my poor english,when I run mmdetection, I encountered a situation,I get different result in same config,I want to ask can you get the same result in mmdetection?
if you can't get the same result,How did the numbers(some like mAP,AP0.5) on the paper come from?Is it training multiple times to get the average?
thank you

Thank you so much for your sharing.But I'm having some problems when I run the second step.
2. Convert a full frame COCO-like dataset to a tiled one

First problem

I fixed this by modifying the following line of code
tiles = split_on_tiles(img, bboxes, **vars(args))

Second problem

How can i solve this problem?Thanks

I haven't run the code yet but as mentioned in the paper, the method perform quite well with counting also. I want to know if counting result has also available in the code yet?

Particularly the one for your final best submission.
Thanks a lot for sharing your great work.

Hello,

Are the model weights available somewhere ?

Thank you

Thank you for the work.
My question is for the tile-combine option, I assume there is a step where the tiled images predicted JSON file labels need to be combined into one JSON file label corresponding to the original image size. Where is that script located?
I am not able to find it.

The soonest response is highly appreciated.
Thank you!

Hello, I finished training and generated the result JSON file. But when evaluating with COCOeval on JSON file and test JSON, the calculation is prolonged, maybe one hour. I want to ask whether you have occurred this problem.