score-zeroshot: A Python repository from hoseong-kim

Semantically Consistent Regularizer (SCoRe)

Statistical Visual Computing Lab (SVCL)

University of California, San Diego

Introduction

This repository contains the source code for "Semantically Consistent Regularization for Zero-Shot Recognition", CVPR, 2017.

Implementation was written by Pedro Morgado. If you encounter any issue when using our code/models, let me know.

Citation

Please cite our paper if it helps your research:

@inproceedings{MorgadoCVPR17,
    author={Pedro Morgado and Nuno Vasconcelos},
    title={Semantically Consistent Regularization for Zero-Shot Recognition},
    booktitle={Computer Vision and Pattern Recognition (CVPR), IEEE Conf.~on},
    year={2017},
    organization={IEEE}
}

Pedro Morgado and Nuno Vasconcelos. 
Semantically consistent regularization for zero-shot recognition.
Computer Vision and Pattern Recognition (CVPR), IEEE Conf. on, 2017.

Prerequisites

python 2.7. (with scikit-image and lmdb)
caffe (with pycaffe)

Tour

Code

score_model.py: Defines the SCoRe model.
score_train.py: Training script.
score_eval.py: Evaluation script.
tools/prepare_LMDBs.py: Script for preparing LMDBs used in score_train.py and score_test.py.

Type python xxx.py --help for usage options.

For a better understanding of the workflow of our code, a script train_eval_CUB.sh is provided for training a SCoRe model on the CUB dataset. This script 1) downloads all required data (CUB and caffemodels for standard CNNs), 2) prepares LMDBs for both images and semantics, 3) trains the classifier and 4) evaluates on source and target (Zero-Shot) classes.

Data

Semantic codewords: Pre-extracted for all classes in both AwA and CUB datasets (see data/).
Partition into source and target classes: see classes.txt, train_classes.txt and test_classes.txt at data/${DB}.
Training sets: data/${DB}/train_images.txt
Test sets: data/${DB}/testRecg_images.txt (source classes) and data/${DB}/testZS_images.txt (target classes).

Results

Mean class accuracy for source and target classes using three architectures: AlexNet, GoogLeNet and VGG19.

Note: Lagrangian coefficients were tuned for Zero-Shot MCA on a set of validation classes. Obtained coefficients are shown below.

Animals with Attributes

Semantics	Source Classes	Target Classes	Semantic Coeff	Codeword Coeff
Attributes	72.5 / 85.1 / 84.6	66.7 / 78.3 / 82.8	0.01	10.0
Hierarchy	74.4 / 84.2 / 84.5	52.3 / 61.2 / 60.7	0.05	1.0
Word2Vec	76.7 / 86.7 / 85.8	51.9 / 60.9 / 57.9	0.01	0.5

Key: AlexNet / GoogLeNet / VGG19

Caltech-UCSD Birds-200-2011

Semantics	Source Classes	Target Classes	Semantic Coeff	Codeword Coeff
Attributes	61.7 / 71.6 / 70.9	48.5 / 58.4 / 59.5	0.01	1.0
Hierarchy	60.2 / 73.1 / 69.6	24.2 / 31.8 / 31.3	0.05	5.0
Word2Vec	61.4 / 73.6 / 71.9	26.0 / 31.5 / 30.1	0.01	1.0

(↓ Reproduced results in Caltech-UCSD Birds-200-2011 dataset)

Semantics	Source Classes	Target Classes	Semantic Coeff	Codeword Coeff
Attributes	62.2 / 73.4 / 69.6	46.7 / 58.5 / 56.3	0.01	1.0
Hierarchy	60.6 / 71.9 / 68.1	24.3 / 25.0 / 27.7	0.05	5.0
Word2Vec	60.7 / 73.4 / 73.9	26.9 / 30.5 / 29.2	0.01	1.0

(↓ Difference between reproduced results (R) and paper ones (P): (R) - (P))

Semantics	Source Classes	Target Classes
Attributes	0.5 / 1.8 / -1.3	-1.8 / 0.1 / -3.2
Hierarchy	0.4 / -1.2 / -1.5	0.1 / -6.8 / -3.6
Word2Vec	-0.7 / -0.2 / 2.0	0.9 / -1.0 / -0.9