KakaoMap Shortens Image Classification Model Development Time Using CLIP Pre-classification and OOD Filtering

When developing the review image classification model, KakaoMap collected approximately 200,000 data points. To reduce the time required for full labeling, it built the training dataset by pre-classifying about 40% of the total data using Zero-shot classification with the CLIP model. ConvNext, known for its structural efficiency, was selected as the main model, and it was confirmed to have superior training and convergence speeds compared to ResNet. Additionally, to address misclassifications caused by the wide distribution of the spam category, post-processing OOD filtering based on Confidence scores (with Temperature scaling applied) was added to strengthen the conservatism and accuracy of spam classification.