Suboptimal results when running landmark detection demo

[soothingni]

Hi,

I am planning to apply parts of the Deep Fashion project to one of my own and had some questions regarding your demo code.

When running the landmark detection demo code, the program returns insatisfactory results even when testing it on provided demo images.

Attached is the result when we ran the code on 07_1_front.jpg.

Is this how the pretrained demo model supposed to operate? And is it normal that the model returns different landmark points every iteration?

Thanks!

[veralauee]

These demo images are used for image retrieval task, not "landmark detection".
We implemented the basic method for landmark prediction task, so it may not achieve high accuracy. In your demo image, it seems just 1 landmark correctly predicted.
We will update the demo images for landmark prediction.
And also improve the performance of landmark prediction in the future.

[HanSeungYeob]

I have a similar question also.
I try to test the custom images for the demo with the zoo model(vgg landmark latest.pth).

But the result is wired. Hard to catch the landmark points.
It looks like every time catch the left upper part of the images.

If the my code is wrong to take the points.

new_image = cv2.circle(new_image, (pred_lm[vis_idx][0],pred_lm[vis_idx][1]), 5, (0,255,0), -1)

Is there any processing code to apply the landmark point to your photos to improve this problem?

[liuziwei7]

The input to the landmark detection model should be cropped clothing bounding box region instead of full image. Hope this could help.

[JYP20050208]

地标检测模型的输入应该是裁剪后的服装边界框区域，而不是完整图像。希望这能有所帮助。

随机预处理RandomResizedCrop和RandomHorizontalFlip破坏了坐标映射关系，即使通过pred_lm[i][0] * (w/224)计算坐标，也无法对应到原始图像中地标的真实位置