I'm working on a project to build a face recognition system and I have a question: do different camera angles affect the performance of the deep learning model?

For example, in CCTV, training data and testing data are taken using different camera heights (assume, for training data, the camera height is 150 cm and, for testing data, the camera height is 200 cm, so the camera angles also can be different). Is this possible to implement? Because when I tested the model with testing data, the results were not that good and it always had false recognition.


1 Answer 1


Yes, changing camera height can affect the performance of models in applications like face recognition. Because it will change the face-capturing angle and size of the face captured. But do not worry, it is a common issue which we face on a regular basis.

To resolve, you need to do some work mentioned below:

  • The main thing for face recognition is face detection. For the face detection model, you need to do augmentations like angles, lighting, resizing, and rotation.

  • You also need to use a model to extract a few points of the face like both eyes, both the start and end points, and nose points. Based on that data, you need to do a warp affine to match the ideal face. Thus, all the data is in normalization format.

  • Choose a neural network like arc face which can work with pose invariant features with angular marginal loss.

Keypoints are augmentation of the data, as the preprocessing step adds warp affine for face normalization and uses a model that can handle pose invariant features.

  • $\begingroup$ Okay, thank you for the explanation. One more question. If I want to build the model for face recognition using this methods, without using another model to extract few point of the faces, just label them manually using LabelImg, is it still possible to do it? @Hiren $\endgroup$
    – Naay
    Dec 13, 2023 at 6:32
  • $\begingroup$ Yes, @Naay you can use model architectures like MTCNN, blazeface or retinaface similar. The method you mentioned is used for Detection only you need to use multi-tasking Vision Models as I mentioned. $\endgroup$ Dec 13, 2023 at 6:39

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .