Few–Shot Action Recognition

  Valeti Likhitha Chowdary,  Nidadavolu Bhavya Sree Ratna,  Ankitha Lunavath,  Rajesh Kandakatla

The goal of few-shot action identification is to detect new action classes using only a few sets of training samples. The majority of available approaches use a meta-learning approach combined with episodic training. The few samples in a meta-training job are divided into support and query sets in each episode. The former is used to construct a classifier, which is subsequently assessed on the latter using a query-centered loss for updating the model. However, there are two key limitations, which are: a lack of data efficiency attributable to the query-centered only loss design and an inability to cope with outlying samples and inter-class distribution overlapping anomalies raised within the support set. We address these shortcomings in this study by introducing a new Prototype-centered Attentive Learning (PAL) model consisting of two revolutionary components. To make maximum use of the minimal training samples in each episode, a prototype-centered contrastive learning loss is introduced to enhance the traditional query-centered learning objective. Second, PAL incorporates a hybrid attentive learning method that can reduce the adverse effects of outliers while also promoting class separation.

Keywords- FSL - Few-Shot Learning PAL - Prototype-centered Attentive Learning

Unique Identification Number - IJEDR2202031
Page Number(s) - 181-185
Pubished in - Volume 10 | Issue 2 | June 2022
DOI (Digital Object Identifier) -    http://doi.one/10.1729/Journal.30748
Publisher - IJEDR (ISSN - 2321-9939)

  Valeti Likhitha Chowdary,  Nidadavolu Bhavya Sree Ratna,  Ankitha Lunavath,  Rajesh Kandakatla,   "Few–Shot Action Recognition", International Journal of Engineering Development and Research (IJEDR), ISSN:2321-9939, Volume.10, Issue 2, pp.181-185, June 2022, Available at :http://www.ijedr.org/papers/IJEDR2202031.pdf