cccorn / ap-loss Goto Github PK

Hi!

glad to meet you. a few days ago, I reviewed your paper and code.

At that time, I found that definition of anchor makes default boxes with real pixel coordinate (i.e, each default box is consist of large scale pixel coordinate).

Why didn't it perform that with normalization coordinate? (e.g, [0.01, 0.01, 0.05, 0.05] means that this box has x_min, y_min, width and height respectively as relative of size of image)

Normally, SSD has used this way.

Thanks!

Hi,

I am planning to use different backbones (resnext-101, resnet-152) different than that you've already provided. As you said in readme, the backbone models are converted from caffe version. Could you please provide the converter tool that you used?

Thank you for your time.
Baris.

Hi,

I appreciate your good work. Read me says that test-dev2017 folder under images should be used. But source code doesn't refer those images. Is it necessary to have test-dev2017 images and what for?

I have read your paper.How do you calculate score?Which formula can you point out?

Hi,

Thank you for the method and the code.
Currently, I am running the code with the default setting (2 GPUs and 8 images/GPU) on 2 V100 GPUs and approximately it will take 7-10 days, which seemed a bit long for me. So, can you specify your training time with your default configuration for me to validate olease?

Kemal

I'm wondering if there is a mistake in line 253 in AP-loss/lib/dataloader/dataloader.py.
Maybe it should be :
valid_boxes.append([x1,y1,x2,y2,cls])

Did I miss anything?

Hi,

I noticed that while in the paper in Equation 5 during AP derivation, you add 1 both to the nominator and to the denominator following simply the formulation of AP.

But in this line of the code, you add 0.5 to the variable a and then estimate the precision by a/(a+b), which implies 0.5 is added to nominator and denominator. I understand all of your code and match with the paper, but I only could not see why you add 0.5 instead of 1 here. Am I missing something?

Thanks for your time

Kemal

Thank your for share great project. Some question not solved and need help. as follow:

1. code1 Why not use for get batch max prec that purpose is get max prec otherwise apply to Interpolation.

# get global max prec in one batch
max_prec = 0
for i in batch:
    current_prec = 1 - L(ij)
    max_prec =max_prec if max_prec>current_prec else current_prec
# same as your code
for ii in order:
        ........
        if (max_prec>=current_prec):
            tmp2*=((1-max_prec)/(1-current_prec))
        valid_bg_grad+=tmp2
        prec[ii]=max_prec 

2. code2, why tmp2 is grad?

Hope your reply.

Hi,

As I see test_img_size parameter in the config file determines the size of the shorter side to be set to test_img_size[0] by ensuring that the longer side does not exceed test_img_size[1], so I think defining a test_img_size=[400,500,600,...] will not evaluate multi-scale testing.

Is there a way to have the results with multi-scale testing or the code does not allow it?

Kemal

Nice work!

But I have a question about the ranking samples, I will be appreciated if you could help me about that.

Are those samples belonging to one ground thruth or just all positive samples in one image, or all positive samples in a batch.

Hi, long time no see. sir!

with congratulating our second meet...

the last time, thank for your reply.
Above, following the code, I have a new question for regression(localization) formulation.

Would you please refer the each line 166, 168, 174 and 176.

why did you limit upper bound of regression_diff_abs by using torch.le()
and if regression_diff_abs don't satisfy the condition, why dose this subtract - 0.5/1.0 or meet torch.sign() ?

Thank you.

Hi, I am reading your repo and some of your code really confused me.

1. Anchor scale
   

   AP-loss/lib/config.py

   Line 16 in 79ba97c

   anchor_scales=np.array([2**0,2**(1.0/2.0)])

   
   is [ 2 ** 0,2 ** (1.0/2.0)], but not [ 2 ** 0, 2 ** (1.0 / 3), 2 ** (2.0 / 3)]

2. Epoch here
   

   AP-loss/lib/config.py

   Line 21 in 79ba97c

   dataset_coco={'dataset':'coco', 'path':'data/coco', 'train_set':'train2017', 'test_set':'val2017', 'epochs':100, 'lr_step':[60,80]}

   
   is 100，but standard coco dataset only trains about 12 epoch

3. Input size here
   

   AP-loss/lib/config.py

   Line 12 in 79ba97c

   train_img_size=512

   
   is 512, not [800, 1333]

could you explain it kindly ?

Just......I'm curious. What are the differences between this work and your 2019 CVPR AP-Loss?