The program runs successfully in practice without any errors; the issue lies with the quality of the landmark detection results.

[wanlipeng]

Hello. I am using the provided pre-trained weights and running the command python singleImage.py -f img. With the default setting args.arch = 'mobilenet_v2', the resulting output images show very poor landmark detection performance, particularly for standard frontal-face images. Is this expected behavior? @choyingw