finetune

Fine-tune pre-trained model in PyTorch

By using our code, we reproduce the results of three popular fine-grained benchmarks.(i.e., Bird, Aircrafts and Cars) We will keep updating the results of this page.

Our experiments are running on

• PyTorch 0.4 or above
• 2 x 1080Ti
• Cuda 9.0 with CuDNN 7.0

Change log

2019/01/22

Add Compact bilinear pooling method (CBP.py).

Update the results of CBP.

2019/11/18

Update the results of fast-MPN-COV-VGG-D. (finetune_mpncovvggd16.sh)

Results (top-1 accuracy rates, %)

All the reproduced results use neither bounding boxes nor part annotations, and the SVM classifier is not performed.m

MPN-COV (ours)

Backbone model	Dim.	Birds		Aircrafts		Cars
		paper	reproduce	paper	reproduce	paper	reproduce
VGG-D	32K	87.2	87.0	90.0	91.7	92.5	93.2
ResNet-50	32K	88.1	88.0	90.0	90.3	92.8	92.3
ResNet-101	32K	88.7	TODO	91.4	TODO	93.3	TODO

Bilinear CNN

Backbone model	Dim.	Birds		Aircrafts		Cars
		paper	reproduce	paper	reproduce	paper	reproduce
VGG-D	262K	84.0	84.0	86.9	86.9	90.6	90.5

Compact bilinear pooling

Backbone model	Dim.	Birds		Aircrafts		Cars
		paper	reproduce	paper	reproduce	paper	reproduce
VGG-D	8K	84.0	83.8	-	TODO	-	TODO

Global average pooling

Backbone model	Dimension	Birds	Aircrafts	Cars
ResNet-152	2K	TODO
ResNet-101	2K
ResNet-50	2K
DenseNet	1K
Inception-v3	2K
VGG-D	4K
AlexNet	4K