i find a bug in submit.py hope you can fix it

[GuanlinLee]

when i use a trained model of FPN,submit.py have an error called

p4_out = torch.add(self.P4_conv1(C4), F.upsample(p5_out, scale_factor=2))
RuntimeError: The size of tensor a (23) must match the size of tensor b (24) at non-singleton dimension 3

i think something wrong when put the pics into model
hope you can fix it

[Xiangyu-CAS]

please change your test size from 368 to 384, because 368 could not be divided by 32. :)

[GuanlinLee]

its already fixed to 384,i tried to find out but it still cant work
btw,im using py3 instead of py2.can this cause the error？

[Xiangyu-CAS]

it is ok to use py3. The error (23 vs 24) is caused by shape miss match in concat. The shape of C4 is 368/16=23, while shape of p5 is 23/2=12 and shape of p5_upsample is 12*2=24. If you have change test size correctly, the exact size of C4 should be 384/16=24.

[Xiangyu-CAS]

so, if this kind of error still occurs on your project, it means you failed to change test size

[Xiangyu-CAS]

Try to replace
util.padRightDownCorner(imageToTest, stride, 128) to
util.padRightDownCorner(imageToTest, 32, 128). I will update this project later.

[Ai-is-light]

@Xiangyu-CAS I have one question about the heatmap. In your project, the code as following,
def _generate_heatmap(img, kpt, stride, sigma):
height, width, _ = img.shape
heatmap = np.zeros((height / stride, width / stride, len(kpt) + 1), dtype=np.float32) # (24 points + background)
height, width, num_point = heatmap.shape
start = stride / 2.0 - 0.5

num = len(kpt)
for i in range(num):
	**if kpt[i][2] == -1:  # not labeled 
		continue**
	x = kpt[i][0]
	y = kpt[i][1]
	for h in range(height):
		for w in range(width):
			xx = start + w * stride
			yy = start + h * stride
			dis = ((xx - x) * (xx - x) + (yy - y) * (yy - y)) / 2.0 / sigma / sigma  # the function of the dis
			if dis > 4.6052:
				continue
			heatmap[h][w][i] += math.exp(-dis)
			if heatmap[h][w][i] > 1:
				heatmap[h][w][i] = 1

**heatmap[:, :, -1] = 1.0 - np.max(heatmap[:, :, :-1], axis=2)  # for background**
return heatmap

I'm confused by getting the heatmap. My question is , if one picture only has 13 keypoints (the number of the total keypoints ), and the batch_size is 1. I want to know which number is the ground truth(or the ground truth heatmap). The number is 24+1 or 13+1 ??
Looking forward any answers.
Thanks.

[Xiangyu-CAS]

@Ai-is-light There are several categories in dataset, 24 + 1 classes is designed for all kinds of pictures. If one picture only have 13 keypoints, the other keypoint heatmaps (24-13) should be blank.

In my solution, I used one model to predict all these 5 categories for convinient, but you could get a more decent performance by training several models for each kind. We have got 7% in leaderboard by ensembling 5 models.

[Ai-is-light]

Good score!
Thanks for sharing your idea, I did that for the baseline.
I have been training 5 models for each class during long time training. But, there are some question .Firstly, I upload my results without following the order of the test.csv, and ; secondly, when I test the dress-class , I directly set the other keypoints which are not included in the dress-class, like the keypoints of trousers set -1_-1_-1. However, I got the answer of the system is invalid upload result.
When you ensemble 5 models, taking the dress class as an example, did you directly set the trousers of -1_-1_-1?
Looking forward any answers.
Thanks.

[Xiangyu-CAS]

@Ai-is-light My test results followed the exact order of test.csv. I am not very clear about your statements, when you test dress-class, you don't need to deal with other classes, because every class has a corresponding model.

[argman]

这比赛你还在做吗

[Xiangyu-CAS]

@argman 当业余爱好在做。。。你是旷视文字组的么。。。

[argman]

@Xiangyu-CAS 不是啊，也是业余选手:)

[Xiangyu-CAS]

@argman 好吧，hh，我去年也做了一段时间文字检测，当时还看了下你的EAST实现，感觉 ID好眼熟。你要也参加的话可以交流交流。。。

[hongge831]

@Xiangyu-CAS modidy.py 是啥用的？

[Xiangyu-CAS]

@hongge831 漏检的惩罚权重较大，modify是为了投机取巧的。如果两个对称点里面漏了一个，那么自动按照中心对称原则计算出另一个点。事实证明用modify了之后，评测性能提升了