BiSPL: Bidirectional Self-Paced Learning for Recognition From Web Data

Deep learning (DL) is inherently subject to the requirement of a large amount well-labeled data, which expensive and time-consuming obtain manually. In order broaden reach DL, leveraging free web data becomes an attractive strategy alleviate issue scarcity. However, directly utilizing collected train deep model ineffective because mixed noisy data. To address such problems, we develop novel bidirectional self-paced (BiSPL) framework reduces effect noise by from in meaningful order. Technically, BiSPL consists two essential steps. Relying on distances defined between samples labeled source samples, first, with short are sampled combined form new training set. Second, based set, both easy hard initially employed models for higher stability, gradually dropped reduce as progresses. By iteratively alternating steps, converge better solution. We mainly focus fine-grained visual classification (FGVC) tasks their corresponding datasets generally small therefore face more significant scarcity problem. Experiments conducted six public FGVC demonstrate that our proposed method outperforms state-of-the-art approaches. Especially, suffices achieve highest stable performance when scale set decreases dramatically.