class_list.txt

[omg777]

Hi, thanks for your nice codes. I modified class_list.txt and then run main.py, It shows aeroplane AP, bicycle AP, bird AP ,... regardless of the class_list file. How can I change this?

[Cartucho]

@omg777 Hello!

So you want to change the name of one of the classes? For that, you should use the code extra//rename_class.py

[omg777]

@Cartucho Thanks for your quick reply. my class_list is “sheep”, “boat” , “horse” but in result plot, there are 20 classes. I want to draw a plot with my own class_list.

[Cartucho]

Hey, do you think you could send me a sample of your files to here: [email protected]

This way I could test the code in the same conditions! :+1:

[omg777]

@Cartucho I solved this problem thanks. but the results from your code are quite different from other mAP codes.(marvis pytorch yolov2 and alexeyAB yolov2) There was too much mAP gap, so I tried to test. I tested only one bbox (cat class), I set GT(bbox: 13 325 84 374) and predicted (:cat 0.026235 49 325 84 374)

In your codes, the prediction got mAP 100% and the other codes got 0.00%

I want to know the differences between your codes and others. thanks!

[Cartucho]

Hello,

I tested only one bbox (cat class), I set GT(bbox: 13 325 84 374) and predicted (:cat 0.026235 49 325 84 374)

In your test, the prediction and GT seem to match! So the mAP should be 100% (given that you only have 1 prediction and it was correct). Don't you agree?

[omg777]

@Cartucho Actually, I dont know exactly how to calculate mAP. AlexeyAB yolo codes are quite realiable because many people have used this and no mentioned problem. Im just wondering why results of mAP are different. Thanks!

[omg777]

@Cartucho Sorry to border you :) In default settings, difficult samples in PASCAL are involved to get mAP or ignored?

[Cartucho]

Are you sure you are using AlexeyAB's code correctly?

Yes, the difficult samples will be ignored but you need to define which ones are difficult.

[omg777]

@Cartucho I tested on mAP from here ([https://github.com/marvis/pytorch-yolo2]) mAP code is same with AlexeyAB`s. Can you check the differences compared your codes?

[Cartucho]

This repo is doing the same as the official Pascal VOC's code for evaluating the neural nets during the previous competition.

I will have a look at their codes during this weekend.

[Cartucho]

@omg777 I had a look at this code.

Just to make sure did you ever got a result different than the 0%? You could be getting 0% due to some wrong formatting in the used files.