Baseline Results for Fine-grained Visual Recognition
DESCRIPTIONS
PyTorch code for baseline results for fine-grained visual recognition.
Fine-grained recognition tasks aim at identifying the specific species of
a bird, or the model of an aircraft. It is a challenging task because the
visual differences between the categories are subtle and can be easily
overwhelmed by the change of pose, viewpoint, or background.
Commonly used fine-grained visual recognition datasets.
---------------------------------------------------
Dataset | # classes | # train | # test
---------------------------------------------------
CUB | 200 | 5,994 | 5,794
Cars | 196 | 8,144 | 8,041
Aircraft | 100 | 6,667 | 3,333
---------------------------------------------------
ResNet-50 and VGG-16 are two mostly widely used backbone networks for
fine-grained visual recognition. These two networks are pre-trained on
ImageNet.
Accuracy of ImageNet pre-trained models.
-----------------------------
Model | ImageNet (%)
-----------------------------
VGG-16 | 71.59
ResNet-18 | 69.76
ResNet-50 | 76.15
-----------------------------
For fine-grained fine-tuning, only the category label is used. The input
images are resized resized to a fixed size of 512*512 and randomly cropped
into 448*448. Random horizontal flip are applied for data augmentation.
For VGG-16, two-step fine-tuning (fine-tune FC layer and then fine-tune all
layers) is essential for several bilinear pooling based methods. However,
two-step fine-tuning does not much help baseline results very much.
Test accuarcy of baseline results.
----------------------------------------
Input size | Model | CUB (%)
----------------------------------------
224 | VGG-16 | 74.47
448 | VGG-16 | 78.79
----------------------------------------
224 | ResNet-18 | 76.89
224 | ResNet-50 | 80.64
448 | ResNet-18 | 82.26
448 | ResNet-50 | 85.99
----------------------------------------
PREREQUIREMENTS
Python 3.7, where Anaconda is preferred.
https://www.anaconda.com/distribution/#download-section
PyTorch 1.0.0
https://pytorch.org/
Visdom (optional)
https://github.com/facebookresearch/visdom/
USAGE
Prepraration
$ mkdir model; cd src/
Modify line 282 of train.py to be your custom data path.
ResNet model fine-tuning
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --arch resnet50 \
--input_size 448
VGG model fine-tuning
$ CUDA_VISIBLE_DEVICES=0,1,2,3 python train.py --arch vgg16 \
--input_size 448
AUTHOR
Hao Zhang (zhangh0214@gmail.com)
LICENSE
CC BY-SA 4.0