Kasteev Museum Models and Dataset Iteration 1 & Iteration 2

The first iteration of the dataset of approximately 60 labeled images per class was created. The number of classes is 39. The second iteration of the dataset has limited sample size that led to insufficient recognition performance (~30% accuracy during in-museum testing). To address this, the dataset was augmented by repeating the same video processing methodology, resulting in an average of 120 images per class. This enhancement allowed the models to generalize better and significantly improved recognition accuracy (up to 96%).

本数据集初代版本每类别包含约60张标注图像，共计39个类别。该数据集第二版样本量有限，导致识别性能不足，馆内测试时准确率仅约30%。为解决该问题，通过复用同一套视频处理流程对数据集进行增强，最终使每类别平均图像数量提升至120张。此次数据增强有效提升了模型泛化能力，识别准确率亦得到显著改善，最高可达96%。