intra-Modal Constraint Loss For Image-Text Retrieval

This paper describes the approach for the Parasitic Egg Detection and Classification in Microscopic Images challenge. Our solution relies on a robust deep learning pipeline implementing a five-fold training schema to pursue the challenge goal. The final methodology utilizes the TOOD model, further enhanced by slicing-aided fine-tuning and inference. The slicing helps to overcome the image size invariability of the dataset and allows the model to access all images in high resolution, and consequently helps it learn detailed features needed to distinguish different classes and find a precise object position. Our results demonstrate the importance of proper data analysis and consequent pre and post-processing to improve prediction performance.