Improve noise estimation performance

🔹 1. Denoising with Confidence Estimation (Failed)

The confidence might be inherently learned during the noise estimation

• Train the noise estimator to also output a confidence map.
• Use this to weight how much noise to subtract:
  denoised = bev_feat - confidence * noise
• Helps avoid over-subtraction in uncertain regions.

---

🔹 2. Feature Consistency Loss : Alrady being applied

• Add a loss between clean BEV features (from the autoencoder) and denoised features (after noise subtraction).
• Encourages your model to recover clean structure.

---

🔹 3. Multi-scale Noise Estimation

unable to apply because the BEV featmap resolution issue

• Use a multi-scale U-Net to estimate noise at different resolutions.
• Subtract noise progressively at each scale.

---

🔹 4. Adversarial Training

• Use a discriminator to distinguish clean vs. noisy BEV features.
• Train the noise estimator + VT encoder to fool the discriminator.
• Similar to denoising GANs.

---

🔹 5. Joint Training with Reconstruction Head : This is the same as the decoder training part

• Add an auxiliary task to reconstruct the input image or GT BEV map from the denoised BEV feature.
• Forces the model to retain meaningful signal while reducing noise.