The study introduces two frameworks, the Progressive Multi-task Anti-noise Learning (PMAL) and Progressive Multi-task Distilling (PMD), to tackle the intra-class variation issue in Fine-grained Vehicle Recognition (FGVR) caused by image noise. PMAL improves recognition accuracy by treating image denoising as an additional task, and PMD transfers the knowledge of the PMAL-trained model into the original network, producing a model with similar accuracy but without additional overheads. These frameworks outperform previous methods on several FGVR datasets.
Publication date: 26 Jan 2024
Project Page: https://github.com/Dichao-Liu/Anti-noise FGVR
Paper: https://arxiv.org/pdf/2401.14336
