4th, Track1 - Image classification for different lighting conditions and styles

2025 IEEE LOW-POWER COMPUTER VISION CHALLENGE, 2025 CVPR Workshop, 2025

Internet

Team: VIP (길다영, 이정윤)

Datasets

  • COCO, OIDv6, Stable Diffusion

Knowledge Distillation

  • student = convnext_tiny.in12k
  • teacher = convnextv2_huge.fcmae_ft_in22k_in1k_512

Freezing : This enhances the effect of transfer learning.

  • freeze_layers : the early part of the model (stem and first stage) is frozen while the head and upper layers are reinitialized, stabilizing the lower layers.
  • progressive_unfreeze : gradually unfreezes layers during training after initially keeping two stages frozen.

Classifier Head

  • A complex head structure improves performance, especially in challenging classification tasks requiring fine-grained distinctions. Different learning rates are applied to the backbone and head, with the head using a 10x higher learning rate to focus on improving classification performance.

Hyperparameter

  • LR = 1e-5
  • kd_temp = 4.0, kd_alpha = 0.7

Results

  • Complie Job Id : j572wzdvp
  • Performace Score : 0.9460580913
  • Time(Microsec) : 1877
  • Accuracy : 0.9460580913