search

ReconAI / AMTrafficPhase2-face-licenseplate-detection

Detection of faces and license plates with various methods
0 stars 0 forks source link

readme

AMTrafficPhase2-face-licenseplate-detection

Task # 2 from Specifications

Prepare and develop end-to-end pipeline (from dataset aggregation to network deployment on end device) for license plate and face detection light-weight neural network. Project must contain a comparison of following approaches:

  1. Detection using miscellaneous Haar-cascade classifier (Viola-Jones detector) implementations (OpenCV, Dlib and other open source)
  2. Detection using Nvidia DetectNet (Resnet-10,18 etc.). DIGITS training and Nvidia TensorRT convertor detection can be considered as an approach.
  3. Other approaches found during a research

I used the following approaches for the project:

  1. Cascade classifier
  2. Lightweight face detector
  3. Detectnet_v2
  4. Retinanet

example testing images for face and license plate detection are also included, for consistency of evaluation across methods.

Method Result Comment
Cascade classifier 25 FPS LP- Mediocre accuracy, but misses small LPs. Faces- poor acccuracy
Lightweight face detector 40 FPS Good speed/accuracy ratio but heavy on resources
Detectnet_v2 28 FPS Good accuracy (0.7 mAP), sufficient speed. Selected solution
Retinanet not measured Failed to train well enough to test further

Detailed specification can be found in each subproject's README

The final submitted solution was built using Detectnet based on Resnet10, demonstrating 28 FPS on Jetson Nano and 0.7 mAP on validation set.