Skip to content

Latest commit

 

History

History
33 lines (25 loc) · 1.46 KB

README.md

File metadata and controls

33 lines (25 loc) · 1.46 KB
Raw

Roadway Intel

Vehicle make and model classification in Tensorflow by fine-tuning VGG16

Dataset

Training dataset consisted of 841 vehicle make/models from CompCars dataset[1]

Architecture

A VGG16 model pre-trained on ImageNet was fine-tuned with CompCars dataset (~100,000 images - 120 images/class)

Training and evaluation pipeline was inspired from Tensorflow's CIFAR tutorial. VGG16 model weights from Caffe Model Zoo were converted to numpy native fromat using caffe-tensorflow tool.

Example Usage

Make sure you've extracted CompCars dataset to PROJECT_ROOT/.

Training:

python -B main.py -t --batch_size 40 --epochs 200 --base_learning_rate=0.0001
--decay_factor=0.1 --decay_epochs=100 --no_gpus=2

Evaluation:

python -B main.py -e --batch_size 40

Results

Accuracy: 93.12% top-5 after 115 epochs Conv5 (Conv5_1, Conv5_2, Conv5_3), FC6, FC7 and FC8 were fine-tuned with vanilla SGD.
Base learning rate of 0.0001 and batch size of 70 were used.

References

[1] Linjie Yang, Ping Luo, Chen Change Loy, Xiaoou Tang. A Large-Scale Car Dataset for Fine-Grained Categorization and Verification, In Computer Vision and Pattern Recognition (CVPR), 2015.