Project 2 Summary

Grades

Overall the class did an excellent job!

28/34 were 19 or higher!

  • 20+ (14 projects): 21.5, 21, 21, (20, 20), (20, 20), 20, 20, 20, (20, 20), 20, 20, (20, 20), 20, 20

  • 19+ (7 projects): (19.5, 19.5), (19.5, 19.5), (19.5, 19.5), 19.5, 19.5, 19, 19,

  • 18+ (2 projects): (18.5, 18.5), 18

  • 17+ (1 projects): 17

Leader Board

  • 1st place, +2 bonus: 1.00 accuracy

  • 2nd place, +1 bonus: 0.9942 accuracy —- Two groups!

  • 3rd place: 0.9883 accuracy

  • 4th place: 0.9766 accuracy —- Three groups!

  • 5th place: 0.9707 accuracy —- Two groups

General Comments:

  1. Insufficient number of epochs

  2. Docker image was not public on Docker Hub

  3. README did not provide enough or clear instruction

  4. Late submission significantly affected the score.

Overview of the 1st-Place Model

Regarding the 1st-place model in this project, the group not only explored the models required by the assignment but also tested a Swin Transformer (Swin-T), a vision transformer architecture. Unlike CNNs, which capture only local spatial patterns through convolutions, Swin-T leverages hierarchical self-attention to learn both local and global image dependencies, enabling significantly stronger feature representation and higher accuracy.