CNN-based off-the-shelf features have shown themselves as a good baseline for trademark retrieval. However, in recent years, the computer vision area was transitioning from CNNs to a new architecture – Vision Transformer. In this paper, we investigate the performance of off-the-shelf features extracted with vision transformers and explore the effects of pre, post-processing, and pre-training on big datasets. We propose a method of joint usage of global and local features, which leverages the best aspects of both approaches. Experimental results on METU Trademark Dataset show that off-the-shelf features extracted with ViT-based models outperform off-the-shelf features from CNN-based models. The proposed method achieves the mAP value of 31.23, surpassing previous state-of-the-art results. We assume that the proposed approach for the trademark similarity evaluation will allow one to improve the protection of such data with the help of artificial intelligence methods. Moreover, this approach will allow one to identify cases of unfair use of such data and form an evidence base for litigation.

The origin of the trademark similarity analysis problem lies within the legal area, specifically the protection of intellectual property. One of the possible technical solutions for this issue is the trademark similarity evaluation pipeline based on the content-based image retrieval approach. CNN-based off-the-shelf features have shown themselves as a good baseline for trademark retrieval. However, in recent years, the computer vision area has been transitioning from CNNs to a new architecture, namely, Vision Transformer. In this paper, we investigate the performance of off-the-shelf features extracted with vision transformers and explore the effects of pre-, post-processing, and pre-training on big datasets. We propose the enhancement of the trademark similarity evaluation pipeline by joint usage of global and local features, which leverages the best aspects of both approaches. Experimental results on the METU Trademark Dataset show that off-the-shelf features extracted with ViT-based models outperform off-the-shelf features from CNN-based models. The proposed method achieves a mAP value of 31.23, surpassing previous state-of-the-art results. We assume that the usage of an enhanced trademark similarity evaluation pipeline allows for the improvement of the protection of intellectual property with the help of artificial intelligence methods. Moreover, this approach enables one to identify cases of unfair use of such data and form an evidence base for litigation.

This paper explores several existing gaming platforms for training and testing artificial intelligence. Application programming interfaces of these platforms are analyzed in order to discover potential complications of porting intellectual systems between them. An approach to unification of programming interfaces in order to overcome these complications is presented. Advantages and potential side effects of this approach are described.