Test image have margin can improve accuracy！
mayidu opened this issue · 4 comments
Hello!
I used megaage asian to train model, and use my images to test. My images was detected faces with MTCNN and saved with margin! When I use the trained model to test the gender with MTCNN detecting face again, and the results bad, but use images with margin (don't detect face with MTCNN), the results very good!
Whether it is related to the training set containing boundaries？
Is it reasonable for me to test like this?
Thank you!
Hello. According to the previous paper, the detected face image should contain extra 40% boundary which is larger than the original detected region. In our experiments, we follow such settings with dlib face detector. MTCNN is also used in our demo code, we also use such setting. However, boundary margin could potentially cause some problem. For example, long hair could be used to detect gender instead of faces, it is not good for the gender task. It still depends on the dataset you use.
非常感谢您的回复！我采用IMDB数据集进行模型的训练，对IMDB数据集进行了一定的处理，使得只包含有人脸区域，而megaage数据集没有给出人脸的位置信息，所以我采用上述的IMDB训练出来的模型作为预训练模型，再在megaage数据集上进行训练的，结果是要比之前有很大的提升，但依然有一些将女性估计为男性。我认为在性别估计方面，发型还是挺重要的一个因素，对于有些人来说，如果没有发型的话，人眼有时候也是很难进行区分的。尤其在我进行megaage数据集性别属性标注的时候，我发现这样的情况还是很多的！
感謝，由於本project畢竟不是為了gender所設計，可能會有很多細節跟以前的性別辨識有所不同，可能需一些調整。
非常感谢！用于年龄估计的话，效果还是很不错的！