The image2label package implements SVM based image concept recognition, with

• a high-level python interface for loading labels, feature vectors, and SVM models, training and applying models
• linear SVMs and fast intersection kernel SVMs
• model compression that allows new models to be added into an existing model
• cross-platform support (linux, mac, windows)

It has been used in multiple tasks including visual categorization [1], tag relevance learning [2], and object localization [3].

##Prerequisites

• The package does not include any visual feature extractors. Features of training and test data need to be pre-computed, and converted to required binary format using txt2bin.py.
• To minimize one's coding effort, the package requires training data and test data to be organized in a fixed structure, see the sample data.

##Getting started

####Setup

• Add simpleknn and image2label to PYTHONPATH
• Change ROOT_PATH in basic/constant.py to local folder where training and test data are stored
• Download sample data (from PASCAL VOC 2008), using the script download_voc_data.sh or via google drive
• Check if everything can be properly loaded by running predict_and_eval.py

The following code samples show how to learn a linear classifier for a given label, say dog, from voc2008train, and evaluate the classifier on voc2008val, with Average Precision (AP) as the performance metric.

####Step 1. Load annotations

from basic.constant import ROOT_PATH
from basic.annotationtable import readAnnotationsFrom
    
rootpath = ROOT_PATH
trainCollection = 'voc2008train'
trainAnnotationName = 'conceptsvoc2008train.txt'
concept = 'dog'
names,labels = readAnnotationsFrom(trainCollection, trainAnnotationName, concept, skip_0=True, rootpath=rootpath)
name2label = dict(zip(names,labels))

####Step 2. Load feature vectors

from simpleknn.bigfile import BigFile

feature = 'dsift'
train_feat_file = BigFile(os.path.join(rootpath,trainCollection,'FeatureData',feature))
renamed,vectors = train_feat_file.read(names)
Ys = [name2label[x] for x in renamed]

####Step 3. Train model

from fastlinear.liblinear193.python.liblinearutil import train as train_model

svm_params += ' -s 2 -B -1 '
model = train_model(Ys, vectors, svm_params + ' -q')

####Step 4. Compress trained model for fast prediction

from fastlinear.fastlinear import liblinear_to_fastlinear as compress_model

feat_dim = train_feat_file.ndims
new_model = compress_model([model], [1.0], feat_dim, params={})

####Step 5. Prediction

testCollection = 'voc2008val'
testAnnotationName = 'conceptsvoc2008val.txt'

from basic.util import readImageSet
test_imset = readImageSet(testCollection, testCollection, rootpath=rootpath)
test_feat_file = BigFile(os.path.join(rootpath,testCollection,'FeatureData',feature))
test_renamed, test_vectors = test_feat_file.read(test_imset)
ranklist = [(test_renamed[i], new_model.predict(test_vectors[i])) for i in range(len(test_renamed))]
ranklist.sort(key=lambda v:v[1], reverse=True)

####Step 6. Evaluation

names,labels = readAnnotationsFrom(testCollection, testAnnotationName, concept, skip_0=True, rootpath=rootpath)
test_name2label = dict(zip(names,labels))
sorted_labels = [test_name2label[x[0]] for x in ranklist]

from basic.metric import getScorer
scorer = getScorer('AP')
perf = scorer.score(sorted_labels)

See predict_and_eval.py for a complete code to compute AP scores for an array of concept models.

####Save model to file # linear SVMs from fastlinear.fastlinear import fastlinear_save_model as save_model

# fik SVMs
from fiksvm.fiksvm import fiksvm_save_model as save_model

save_model(model_file_name, new_model)

####Load model from file

# linear SVMs
from fastlinear.fastlinear import fastlinear_load_model as load_model

# fik SVMs
from fiksvm.fiksvm import fiksvm_load_model as load_model

model = load_model(model_file_name)    

####Add a new (compressed) model to an existing model

# model_1 <- model_1 * w1 + model_2 * w2
model_1.add_fastsvm(model_2, w1, w2)

[1] Xirong Li, Cees G. M. Snoek, Marcel Worring, Dennis Koelma, Arnold W. M. Smeulders, Bootstrapping visual categorization with relevant negatives, IEEE Transactions on Multimedia, 15(4):933-945, 2013

[2] Xirong Li, Cees G. M. Snoek, Classifying tag relevance with relevant positive and negative examples, ACM Multimedia, 2013

[3] Xirong Li, Qin Jin, Shuai Liao, Junwei Liang, Xixi He, Yujia Huo, Weiyu Lan, Bin Xiao, Yanxiong Lu, Jieping Xu: RUC-Tencent at ImageCLEF 2015: Concept Detection, Localization and Sentence Generation, CLEF Working Notes 2015