This is an implementation of the algorithm described in the following paper:
I.A. Kalinovskii, V.G. Spitsyn,
Compact Convolutional Neural Network Cascade for Face Detection,
http://arxiv.org/abs/1508.01292
This code used only for evaluation of face detectors on FDDB, AFW and IJB-A benchmarks or video data. It does not contain CNN detector. To view the test results it is necessary extract files from rar-archives to the root folder (Data/FDDB_TEST) and run the Matlab script.
If you use the provided code/binaries or data for your work, please cite this paper.
Demo video:
The precision detection can be increased by combination detectors.
№ | method | true | false | recall | precision | mean time, ms | min time, ms | max time, ms | demo |
1 | Viola–Jones1 | 18679 | 3587 | 46.7% | 83.9% | 118 | 79 | 136 | HD |
2 | Compact CNN | 18788 | 908 | 47.0% | 95.4% | 12 | 10 | 17 | HD |
3 | Compact CNN (optimized for precision) |
16627 | 379 | 41.6% | 97.8% | 12 | 10 | 17 | |
4 | Hybrid2 (2 + 1) | 16684 | 93 | 41.7% | 99.4% | 17 | 10 | 800 | HD |
2without code optimization
Running on GPU Nvidia GeForce GT 640M
CPU Intel Core i7-3610QM (2.3GHz, without Turbo Boost)
Search settings: minSize – 40×40, scaleFactor – 1.2, minNeighbors – 2, without tracking
For any additional information contact me at kua_21@mail.ru.
Copyright (c) 2015, Ilya Kalinovskii. All rights reserved.