Abstract
Most of the existing algorithms for zero-shot classification problemstypically rely on the attribute-based semantic relations among categories torealize the classification of novel categories without observing any of theirinstances. However, training the zero-shot classification models still requiresattribute labeling for each class (or even instance) in the training dataset,which is also expensive. To this end, in this paper, we bring up a new problemscenario: "Can we derive zero-shot learning for novel attributedetectors/classifiers and use them to automatically annotate the dataset forlabeling efficiency?". Basically, given only a small set of detectors that arelearned to recognize some manually annotated attributes (i.e., the seenattributes), we aim to synthesize the detectors of novel attributes in azero-shot learning manner. Our proposed method, Zero-Shot Learning forAttributes (ZSLA), which is the first of its kind to the best of our knowledge,tackles this new research problem by applying the set operations to firstdecompose the seen attributes into their basic attributes and then recombinethese basic attributes into the novel ones. Extensive experiments are conductedto verify the capacity of our synthesized detectors for accurately capturingthe semantics of the novel attributes and show their superior performance interms of detection and localization compared to other baseline approaches.Moreover, we demonstrate the application of automatic annotation using oursynthesized detectors on Caltech-UCSD Birds-200-2011 dataset. Variousgeneralized zero-shot classification algorithms trained upon the datasetre-annotated by ZSLA show comparable performance with those trained with themanual ground-truth annotations. Please refer to our project page for sourcecode: https://yuhsuanli.github.io/ZSLA/