Negative Samples about libfacedetection.train

[ShiqiYu]

Use the score to filter out

[PhenomenalOnee]

Sir the confidence is high for the false detection, is there a way to add negative samples in training

[PhenomenalOnee]

Also increasing the confidence decreases the detection rate too

[ShiqiYu]

you can show some detetection results with the confidence scores.

[vladkol]

I noticed this when detection resolution is higher than the network was trained with. When resizing images to 320 before detection (and priors are made with that as well), I'm getting expected results.

[PhenomenalOnee]

@ShiqiYu Sir I have trained the model (original caffe yunetv1) on wider dataset, (I have modified the pytorch model to match the original caffe v1 model), and in this model testing I am experiencing false positives, may be the model is too simple to handle such complex cases, that's why i am asking if you can help in adding negative samples. I am thinking of adding hand images with bounding box and class label as background in training data but getting error as
RuntimeError: invalid argument 2: non-empty vector or matrix expected at /pytorch/aten/src/THCUNN/generic/ClassNLLCriterion.cu:31
how can i modify the multibox loss function in this case to support images only with background label

[PhenomenalOnee]

This image has detection on hand with 86% confidence

[ShiqiYu]

As suggested by @vladkol , you can resize the image to a smaller one, then detect it.

[PhenomenalOnee]

i am doing detection at 160x120 dimension, i just want to know the way i can modify the multiloss function to accept negative samples

[boarder8383]

Got the exact same issue here. I've already resized to 320 and still got tons of false positives, like the back of the head dog faces etc. It will be very helpful to have negative examples as training inputs.

[ShiqiYu]

Can you provide the image to test and debug?

…

On Sun, Sep 6, 2020 at 11:52 PM boarder8383 ***@***.***> wrote: Got the exact same issue here. I've already resized to 320 and still got tons of false positives, like the back of the head etc. It will be very helpful to have negative examples as training inputs. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#16 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABWR4HISG6STMBVQFV554YLSEOV2ZANCNFSM4NSPJI7A> .

[boarder8383]

Need to collect again. I do see hard negative examples code inside multibox_loss.py. Will you be able to share a version that takes negative examples inputs ?

Can you provide the image to test and debug?
…
On Sun, Sep 6, 2020 at 11:52 PM boarder8383 @.***> wrote: Got the exact same issue here. I've already resized to 320 and still got tons of false positives, like the back of the head etc. It will be very helpful to have negative examples as training inputs. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#16 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABWR4HISG6STMBVQFV554YLSEOV2ZANCNFSM4NSPJI7A .

Need to

[PhenomenalOnee]

So We need to do some changes to train with negative samples in data.py and multibox_loss.py

In file data.py
In class PreProc , if we pass an image with label index == 0, we can return empty numpy,
if int(labels_t[0]) == 0:
return image_t, np.empty(0)
In the xml file, you can pass background as label for negative images

In multiloss.py
In line where we parse the batches, we need to compare the size of the target
If empty array is encountered we can initialize 0 valued arrays
for idx in range(num):
if len(targets[idx]) < 1 :
#print(targets[idx])
#print("Amazing")
truths = torch.zeros(1,4)
truths = truths.cuda()
labels = torch.zeros(1,1)
labels = labels.cuda()
else :
truths = targets[idx][:, 0:4].data
labels = targets[idx][:, -1].data

You can also perform this initialization of 0-valued arrays in data.py.

Now comes the part what if all items in a batch contains negative samples, than we will have error in calcualting crossentropy because we are calculating neg_indexes based on pos_indexes and if pos_index is 0 we get 0 neg_indexes and empty array is passed in crossentropy
To pass that , i have initialized a tensor when i get this situation like

num_pos = pos.long().sum(1, keepdim=True)
num_neg = torch.clamp(self.negpos_ratio*num_pos, max=pos.size(1)-1)
#######ERRO HANDLE#########
if num_neg.sum().item() == 0.0:
print('++++++++++++++++++++++++++EMPTY',num_neg.size(),num_neg)
p = torch.tensor(np.array([[100]]),dtype=torch.int64)
num_neg = torch.clamp(p.expand_as(num_pos), max=pos.size(1)-1)
num_neg = num_neg.to('cuda:0')
##########################