Uncertainty-42/HGA_Co2SAM_based

HGA Pretrained Checkpoints (PASCAL VOC 2012 & MS COCO 2014)

This repository hosts the official best-performing model checkpoints for the HGA (Hierarchical-Geometric Alignment) weakly supervised semantic segmentation (WSSS) paradigm.

Both checkpoints leverage the high-performance EfficientViT-SAM-XL0 backbone, but employ different decoder architectures optimized for their respective dataset scales.

📊 Benchmark Checklist

Checkpoint Filename	Target Dataset	Backbone	Decoder Type	Reached mIoU
best_model_voc.pth	PASCAL VOC 2012 Val	EfficientViT-SAM-XL0	RC (Resize-Conv) Decoder	84.91%
best_model_coco.pth	MS COCO 2014 Val	EfficientViT-SAM-XL0	CT (Transpose-Conv) Decoder	59.31%

---

⚠️ Critical Setup Instructions (Decoder Mismatch Prevention)

To prevent PyTorch state_dict loading errors (such as KeyError or missing key warnings), you must configure the decoder type in your config.py to match the downloaded checkpoint exactly:

1. For PASCAL VOC 2012 (best_model_voc.pth):

Ensure your decoder configuration is set to use the patched Resize-Convolution (RC) Decoder:

# Inside your config.py for VOC evaluation:
decoder_type = "RC"  # Ensure this matches the RC-Decoder setup

2. For MS COCO 2014 (best_model_coco.pth):

Ensure your decoder configuration is set to use the standard Transposed Convolution (CT) Decoder:

# Inside your config.py for COCO evaluation:
decoder_type = "CT"  # Ensure this matches the CT-Decoder setup

---

🚀 Evaluation & Reproducibility Guide

The evaluation scripts are designed to automatically reconstruct the correct network architecture based on your configuration files.

1. Download the target checkpoint file (.pth) and place it under your local checkpoints/ folder.
2. Clone our official codebase and configure your directory paths: 👉 Uncertainty-42/HGA GitHub Repository
3. Run the evaluation shell scripts directly from the repository root:

# To evaluate PASCAL VOC:
bash val_voc.sh

# To evaluate MS COCO:
bash val_coco.sh

📝 Citation

Our paper is currently under review/pre-print preparation. The official BibTeX citation and arXiv link will be updated here as soon as the preprint is publicly released.