Coldswamp/ResNet50-Matting

ResNet-50 Image Matting Baseline

This repository provides a robust image matting baseline leveraging a ResNet-50 backbone, inspired by the training philosophies of AEMatter.

📊 Experimental Results

We evaluate our model using standard matting metrics. The results below demonstrate the performance of the provided checkpoints.

Checkpoint	SAD ↓	MSE ↓	Grad ↓	Conn ↓
model.ckpt	23.82677	0.00427	8.08990	19.02270
model_better.ckpt	23.47848	0.00396	7.80796	18.89092

Note: model_better.ckpt represents the state-of-the-art (SOTA) performance for this baseline configuration.

---

💡 Key Optimization: Eval-Mode Fine-Tuning

The superior performance of model_better.ckpt is attributed to a specific optimization strategy: Fine-tuning the entire network in eval() mode for a final epoch.

Technical Intuition

In many image restoration tasks (such as Super-Resolution and Matting), Batch Normalization (BN) layers can introduce noise if the mini-batch statistics fluctuate significantly during the final stages of convergence.

To achieve a more stable and high-performance result:

1. The Strategy: After standard training, we perform one additional fine-tuning epoch.
2. The Mechanism: The model is switched to .eval() mode, but gradients are not frozen.
3. The Benefit: By keeping the BN layers in evaluation mode, we use the stable, pre-calculated global running statistics instead of batch-specific statistics. This allows the weights to strictly optimize for the loss objective without being affected by batch-level variance.
4. Validation: This approach has been proven effective in various low-level vision benchmarks to refine details and improve global consistency.

---

🛠 Reproducing the Result

To replicate the performance of model_better.ckpt, ensure you run the final fine-tuning step as follows:

# Final Fine-tuning in Eval Mode
model.eval() # Use global BN stats, but keep gradients active
optimizer = torch.optim.Adam(model.parameters(), lr=1e-5) # Use a smaller LR

for images, trimaps, gt in final_refinement_loader:
    optimizer.zero_grad()
    output = model(images, trimaps)
    loss = criterion(output, gt)
    loss.backward()
    optimizer.step()

# Save the refined model
save_checkpoint(model, "model_better.ckpt")

---
license: apache-2.0
---