These are a couple of ideas I think would be worthwhile to pursue.
1. Genetic Algorithms x Computer Vision
Determine if there’s a smart way to determine a network topology for a given computer vision task. Example: Using neuro-evolution of augmenting topologies (paper here), train a convolutional neural network to do face detection/recognition. Note: Adam Cjazka suggested to look into NAS (neural architecture search).
2. Grid-Search Performance Analysis of Face Detectors & Trackers
There are a lot of face detectors, and there are a lot of object trackers. Set up an exhaustive search of the space to find the best performing/most accurate combinations. From there, generate tracklets for an automated ground truther v2.
3. When to Save a Face Encoding
Traditionally, when a face is first found (usually small/LR), that is when the base encoding is generated. For the task of face recognition, all subsequent detections generate encodings which are then checked against this base encoding. Perhaps there exists a better way to save/update encodings as the quality of the face improves. This ties into Haoyu Chen’s work on “identifiabiliy” of a face.
4. Combine Face Recognition with Action Detection
This one was inspired by a heartbreaker of a Game of Thrones clip. It would be interesting to see if it’s possible to determine which characters are good and bad in a scene based on action detection when paired that with face recognition. ‘Bad’ in this case might be characters that exhibit violent behavior.
5. Effect of Surveillance Camera Depression Angle on Facial Recognition
This one is self-explanatory, but definitely needs an overhauled experimental setup.
6. Revisit Inferring Face Location from Body Bounding Box
Results from last experiments did not prove to be actionable, so the project was dropped, but this could be revisited at a later date.