onepass7b_three_method_visualization_val/visualization_summary.md

OnePass-7B visual comparison

Dataset-level diagnostics

Method	Mean sample IoU	IoU < 0.5	IoU >= 0.75
Qwen-init	0.586179	27.88%	29.92%
STAMP-LoRA-init	0.585672	28.43%	30.60%
SEG-grounding	0.589904	27.12%	31.66%

Pairwise changes

Comparison	Mean delta	Improved >0.01	Regressed >0.01	Nearly unchanged
STAMP-LoRA-init - Qwen-init	-0.000507	35.48%	33.97%	30.56%
SEG-grounding - STAMP-LoRA-init	+0.004233	29.06%	19.71%	51.23%

Selected qualitative cases

Index	Category	Qwen-init	STAMP-LoRA-init	SEG-grounding
118	all_methods_fail	0.0000	0.0000	0.0000
121	all_methods_fail	0.0000	0.4044	0.4314
3163	stamp_lora_gain	0.0004	0.6621	0.0000
1045	stamp_lora_gain	0.8401	0.8375	0.8371
808	seg_grounding_gain	0.0000	0.0000	0.0000
3822	seg_grounding_gain	0.0059	0.0049	0.0186
1641	seg_grounding_regression	0.0000	0.0000	0.0000
2162	seg_grounding_regression	0.7193	0.8151	0.6546
105	largest_method_disagreement	0.0011	0.0007	0.0007
1021	largest_method_disagreement	0.8497	0.8622	0.8625
3778	all_methods_succeed	0.9501	0.9503	0.9357
754	all_methods_succeed	0.8923	0.8976	0.8803

Colors in prediction panels: green=TP, red=FP, blue=FN.

The selected cases are post-hoc qualitative diagnostics based on GT IoU; selection is not part of model inference.