KRMayD/COD10K_CAM_5Way_Segmentation_Checkpoints

COD10K CAM Five-Way Segmentation Checkpoints

This repository contains the five OpenCLIP ViT-B-32-quickgelu checkpoints used in one controlled COD10K CAM segmentation comparison.

Checkpoints

File	Training method
baseline_openai_clip_vit_b32_quickgelu.pt	Unmodified OpenAI CLIP baseline
clip_posttrained_explicitneg_2886_seed0_final_model.pt	Native CLIP post-training on positive image-caption pairs
cliprefine_posttrained_explicitneg_2886_seed0_final_model.pt	Native CLIP-Refine post-training on positive image-caption pairs
gmpo_global_ind_explicitneg_2886_seed0_epoch_3.pt	GMPO with one frozen reference-derived IND four-way vector for all samples
gmpo_sample_ind_explicitneg_2886_seed0_epoch_3.pt	GMPO with a frozen reference-derived IND four-way vector per sample

All checkpoints are plain PyTorch/OpenCLIP state dictionaries compatible with the ViT-B-32-quickgelu architecture.

Shared Data and Evaluation Protocol

• Train split: 2,886 COD10K CAM pairs after removing 154 SHA-256-verified MM-CamObj Easy/Hard benchmark overlaps.
• CSV inputs: filename, filename_neg, Caption, Caption_neg.
• Negative image: Stable-Diffusion inpainted target-mask region composited on the original background.
• Caption policy: image-specific base background caption plus explicit animal presence (T+) or absence (T-) sentence.
• Test split: 2,026 COD10K CAM images.
• Prompt: There is a/an {animal} visually blended into its surroundings.
• Saliency parameters: vbeta=1.0, vvar=0.3, vlayer=8, seed 42.
• Metrics: mean DSC and NSD with 2-pixel tolerance.

The exact paths, settings, data checksum, and metrics are included in comparison_protocol.json and metrics_summary.json.

Segmentation Results

Model	DSC	NSD
Baseline CLIP	0.352689	0.375161
CLIP post-training	0.359849	0.383442
CLIP-Refine post-training	0.359409	0.383454
GMPO Global-IND	0.378582	0.402299
GMPO Sample-IND	0.356773	0.379253

Important Comparison Detail

The train split, starting checkpoint, number of epochs, seed, test images, prompt, saliency settings, and metric code are fixed across the comparison. GMPO necessarily uses the full four-tuple. CLIP and CLIP-Refine use only the positive image-caption pair by their native definitions and retain their native optimizer settings; see comparison_protocol.json for the exact values.