Locality-constrained Linear Coding for Image Classificatio
Citations: 3,201 | Paper Published: 13 Jun 2010 | Institutions: University of Illinois at Urbana–Champaign | Authors: Jinjun Wang, Jianchao Yang1, Kai Yu, Fengjun Lv, Thomas S. Huang1, Yihong Gong
The paper proposes a new coding scheme called Locality-constrained Linear Coding (LLC) to replace the traditional VQ coding in SPM for image classification. LLC utilizes locality constraints and max pooling, achieving state-of-the-art performance with linear classifiers on several benchmarks.
Contributions
The contributions of this paper are the introduction of a novel and practical coding scheme called Locality-constrained Linear Coding (LLC) for image classification, an approximated LLC method to allow fast computation, and an incremental training algorithm to construct the codebook used by LLC. The proposed approach performs remarkably better than traditional nonlinear SPM with linear classifiers on several benchmarks while maintaining computational efficiency even with very large codebooks.
Practical Implications
The practical implications of this paper are that the proposed LLC method provides a promising and efficient approach for image classification, outperforming many existing methods while maintaining computational efficiency. The approximated LLC method allows fast computation even with very large codebooks, making it suitable for real-world applications where speed is important. Additionally, the incremental training algorithm enables construction of an optimized codebook using large-scale training descriptors to further improve performance on various datasets.
Methods
The Methods in this paper include the proposed Locality-constrained Linear Coding (LLC) scheme, which utilizes locality constraints and max pooling to generate image feature representations. The approximated LLC method is also introduced for fast computation with large codebooks. Additionally, an incremental training algorithm is employed to construct optimized codebooks using large-scale training descriptors. Two pooling methods are utilized in the SPM layer: VQ codes and SC codes along with LLC codes as a replacement of traditional VQ coding approach.
Data
The dataset used in this paper is the Caltech-101 dataset, which contains 9144 images in 101 classes including animals, vehicles, flowers etc. The number of images per category varies from 31 to 800.
Results
- The proposed LLC method achieves state-of-the-art performance on several benchmarks including Caltech-101, Caltech-256 and Pascal VOC 2007
- On the Caltech-101 dataset, it achieved an overall accuracy of 91.8%.
- On the PASCAL VOC 2007 dataset, it achieved a mAP score of 60.2%, which is higher than many existing methods such as sparse coding-based approaches and traditional SPM with VQ codes.
- Compared to other fast algorithms like VLAD or FV encoding schemes: The approximated LLC method maintains high accuracy while providing significant computational efficiency even with very large codebooks (O(M + K^2)).
Limitations
This paper does not explicitly mention any limitations or drawbacks of the proposed approach. However, like all research papers, there may be potential limitations that were not addressed in this work and could be explored further by future studies. Additionally, while the proposed LLC method achieves state-of-the-art performance on several benchmarks as reported in this paper, it is important to note that no single algorithm can achieve optimal results for every dataset or application scenario. Therefore, researchers should carefully evaluate different methods based on their specific requirements before making a final decision about which one to use for their particular task at hand.