Can Partial Strong Labels Boost Multi-label Object Recognition?

Convolutional neural networks (CNN) have shown great performance as a global representation for object recognition. However, for multi-label images that contain multiple objects from different categories, scales and locations, single CNN features might not be be optimal. To enhance the robustness and discriminative power of CNN features for multi-label object recognition problem, we propose a multi-view multi-instance framework. This framework transforms the multi-label classification problem into a multi-class multi-instance learning problem by extracting object proposals from images. A multi-view pipeline is then applied to generate a two-view representation of each proposal by exploiting two levels of labels in multi-label recognition problem. The proposed framework not only has the flexibility of utilizing both weak and strong labels or just weak labels, but also holds the generalization ability to boost the performance of unseen categories by available strong labels. Our framework is extensively compared with state-of-the-art hand-crafted feature based and CNN based methods on two multi-label benchmark datasets. The experimental results validate the discriminative power and generalization ability of the proposed framework. When combined with a very-deep network, we can achieve state-of-the-art results in both datasets.