[MODEL] Why Class Imbalance is not working with Fine Graded Architecture (Dlaczego Class Imbalance nie dziala z architektura Fine Grained)

[rojberr]

User Story:

As a computer vision engineer working with fine-grained classification,
I want to understand why traditional class imbalance techniques (weighting, oversampling) fail with my architecture,
So that I can either adapt existing solutions or develop custom approaches for fine-grained imbalance problems.

---

Acceptance Criteria:

1. Comparative analysis of imbalance techniques across architectures:
   • Standard CNN vs. Fine-Grained (e.g., Bilinear-CNN, Vision Transformer)
2. Feature space visualization of minority/majority classes
3. Gradient flow analysis during imbalance correction attempts
4. Minimum of 3 fine-grained datasets tested (CUB-200, Stanford Dogs, etc.)
5. Documentation of failure modes with suggested mitigations

---

Definition of Done:

• Analysis notebook showing:
  • Class separation metrics (intra/inter-class variance)
  • Loss landscape differences
  • Attention map comparisons
• CI pipeline with fine-grained imbalance checks
• Architecture-specific recommendations document
• Performance benchmarks before/after fixes

---