iris_detector
How to train iris detector
See how_to_train_iris_detector_with_caffe_ssd.md
Network design
ResNet10-SSD
Using 4 residual modules as base network, then add SSD's extra layers.
Visualization of network structure (tools from ethereon) http://ethereon.github.io/netscope/#/gist/bc73857987941a56bc45bf4c4ae870b0
ResNet10-SSD with half filter number
The structure is same as ResNet10-SSD except its filter number. When filter number is larger than 32, reduce it by half.
Visualization of network structure (tools from ethereon) http://ethereon.github.io/netscope/#/gist/cf4dccec1f9a6c8f3f125000cd7b97f9
MobileNet-SSD
See https://github.com/chuanqi305/MobileNet-SSD
Visualization of network structure (tools from ethereon) http://ethereon.github.io/netscope/#/gist/e1e8c3c3a450f0502ef8ff6547d5dedb
Experiment
Our iris dataset has 12800 training samples and 3200 test samples. Training on GTX1080Ti. Evaluate on Intel i5 CPU and GTX1080Ti GPU.
Speed test ResNet10+SSD(half) is faster than others.
Network | mAP@0.5 | Speed on Intel i5 CPU(ms) | Speed on GTX1080Ti(ms) | Input resolution |
---|---|---|---|---|
ResNet10+SSD | 1.0 | 20 | 13 | 640x480 |
ResNet10+SSD(half) | 1.0 | 10 | 7 | 640x480 |
MobileNet+SSD | - | 27 | 18 | 640x480 |
ResNet10-SSD
Training
When set confidence threshold to 0.5 and set IoU threshold to 0.5, the accuracy is 100%.
Evaluation
On Intel i5 CPU, the average inference time is 20ms. On GTX1080Ti GPU, the average inference time is 13ms.
Intel i5 CPU
GTX1080Ti GPU
ResNet10-SSD with half filter number
Training
When set confidence threshold to 0.5 and set IoU threshold to 0.5, the accuracy is 100%.
Evaluation
On Intel i5 CPU, the average inference time is 10ms. On GTX1080Ti GPU, the average inference time is 7ms.
Intel i5 CPU
evaluate by opencv3.4 Python API:
evaluate by opencv3.4 C++ API:
GTX1080Ti GPU
MobileNet-SSD
Evaluation
We just evaluate mobilenet_300x300_ssd_iter_3000.caffemodel. On Intel i5 CPU, the average inference time is 27ms. On GTX1080Ti GPU, the average inference time is 18ms. The speed is slower than ResNet10-SSD which has high accurracy on iris dataset, so we stop training.
Questions
Please contact timlee.zh@gmail.com