Interpreting DETR with Respect to Object Queries

In recent years, understanding the decision-making processes of neural network models has become increasingly important. This paper is focusing on this concern and addresses this issue on DETR model which is an object detection framework based on transformers. We explore ways to improve the interpretability and efficiency of this model by proposing and applying new methods for representing this model. A key focus of this research is on object queries specifically analyzing their behavior during the model's validation process. We identify and remove less active object queries from the model, subsequently, measuring the impact on the training time and also the model accuracy. By eliminating 10% of these inactive object queries, we reduce validation time by 3.3% and lower the model’s required FLOPS by 36.25 G. The GitHub repository for this project is available at https://github.com/morteza-shahrabi-farahani/Interpreted-DETR-model