Summary of the seminar on March 19, 2024, by the Faculty of Software Engineering

25/03/2024 - 05:53

To exchange knowledge and enhance expertise, on the morning of March 19, 2024, the Faculty of Software Engineering organized an academic seminar in room E7.3 with the theme "ANALYSIS OF FINE-GRAINED COUNTING METHODS FOR MASKED FACE COUNTING: A COMPARATIVE STUDY" with the participation of professors, lecturers within the Faculty, presented by TS. Quan Chi Khanh An, which was successfully conducted.

The seminar presented the problem of counting the number of masked faces in images, which involves counting the number of faces under various crowd densities and distinguishing between masked and unmasked faces, which is often considered as an object detection task (i.e., face detection). The accuracy of counting is limited, especially at high crowd densities, when faces are relatively small, unclear, and observed at different angles. Moreover, labeling the necessary bounding boxes for training object detection methods is costly. We consider counting masked faces as a fine-grained crowd counting task, which is suitable for addressing the aforementioned challenges with density map regression. However, applying fine-grained crowd counting methods to count masked faces is not straightforward. Appropriate strategies need to be defined for both counting and multi-class classification. We compared strategies of various methods and examined their advantages and disadvantages. These strategies include (1) simple regression and combined regression with detection for counting, (2) using class-aware density maps with semantic segmentation maps and class probabilities for classification, and (3) counting with/without depth information. Analysis of seven crowd counting methods on three datasets with a total of approximately 900 thousand annotations demonstrated that image crowdiness affects the accuracy of simple regression and combined regression with detection for counting. Meanwhile, the most effective method for classification is using semantic segmentation maps. Evaluation of the usefulness of using depth data demonstrated the need for a depth map to achieve high counting accuracy. Our findings will be useful for future research.

Below are some summaries of the results and the participating professors and lecturers:

Capture