{ "methods": [ { "label": "Qwen-init", "checkpoint": "/inspire/hdd/global_user/liuxiaotong-253108540242/yanggang/lihao/lh/or/MLLM-SEG/outputs/fair_7b_onepass6_stamp4_e2_v2/onepass7b/onepass_qwen7b.pt", "rows_csv": "/inspire/hdd/global_user/liuxiaotong-253108540242/yanggang/lihao/lh/or/MLLM-SEG/outputs/fair_7b_onepass6_stamp4_e2_v2/onepass7b_eval_val/onepass7b_eval_rows.csv", "initialization": "base_qwen2_vl_without_stamp_weights", "use_seg_grounding": false, "seg_fusion_alpha": 0.0, "selected_inference_seconds": 11.155093079432845 }, { "label": "STAMP-LoRA-init", "checkpoint": "/inspire/hdd/global_user/liuxiaotong-253108540242/yanggang/lihao/lh/or/MLLM-SEG/outputs/onepass7b_stamp_lora_warmstart_e2/onepass_qwen7b.pt", "rows_csv": "/inspire/hdd/global_user/liuxiaotong-253108540242/yanggang/lihao/lh/or/MLLM-SEG/outputs/onepass7b_stamp_lora_warmstart_e2_eval_val/onepass7b_eval_rows.csv", "initialization": "stamp_7b_lora_warm_start", "use_seg_grounding": false, "seg_fusion_alpha": 0.0, "selected_inference_seconds": 7.2885101940482855 }, { "label": "SEG-grounding", "checkpoint": "/inspire/hdd/global_user/liuxiaotong-253108540242/yanggang/lihao/lh/or/MLLM-SEG/outputs/onepass7b_stamp_lora_seg_grounding_e2/onepass_qwen7b.pt", "rows_csv": "/inspire/hdd/global_user/liuxiaotong-253108540242/yanggang/lihao/lh/or/MLLM-SEG/outputs/onepass7b_seg_grounding_eval_val/onepass7b_eval_rows.csv", "initialization": "onepass_checkpoint_seg_grounding_finetune", "use_seg_grounding": true, "seg_fusion_alpha": -0.001273852656595409, "selected_inference_seconds": 7.553720632568002 } ], "distribution": { "common_samples": 4896, "runs": [ { "label": "Qwen-init", "mean_sample_iou": 0.5861785892137817, "iou_below_0.5_fraction": 0.27879901960784315, "iou_above_0.75_fraction": 0.2992238562091503 }, { "label": "STAMP-LoRA-init", "mean_sample_iou": 0.5856717786875781, "iou_below_0.5_fraction": 0.28431372549019607, "iou_above_0.75_fraction": 0.3059640522875817 }, { "label": "SEG-grounding", "mean_sample_iou": 0.5899042908731187, "iou_below_0.5_fraction": 0.27124183006535946, "iou_above_0.75_fraction": 0.3165849673202614 } ], "pairwise": [ { "comparison": "STAMP-LoRA-init - Qwen-init", "mean_delta_iou": -0.0005068105262035741, "improved_by_more_than_0.01_fraction": 0.3547794117647059, "regressed_by_more_than_0.01_fraction": 0.3396650326797386, "nearly_unchanged_fraction": 0.3055555555555556 }, { "comparison": "SEG-grounding - STAMP-LoRA-init", "mean_delta_iou": 0.0042325121855405715, "improved_by_more_than_0.01_fraction": 0.29064542483660133, "regressed_by_more_than_0.01_fraction": 0.19709967320261437, "nearly_unchanged_fraction": 0.5122549019607843 } ] }, "selected_cases": [ { "index": 118, "category": "all_methods_fail", "instruction": "Please segment the object this sentence describes: \"a black motor bike with two head lights\".", "ious": { "Qwen-init": 0.0, "STAMP-LoRA-init": 0.0, "SEG-grounding": 0.0 } }, { "index": 121, "category": "all_methods_fail", "instruction": "Please segment the object this sentence describes: \"the chair behind the guy wearing the stripes\".", "ious": { "Qwen-init": 0.0, "STAMP-LoRA-init": 0.4044166946027489, "SEG-grounding": 0.43144398837799197 } }, { "index": 3163, "category": "stamp_lora_gain", "instruction": "Please segment the object this sentence describes: \"a woman in an apron and uniform holding two pots\".", "ious": { "Qwen-init": 0.0004290519667742157, "STAMP-LoRA-init": 0.6620860738008585, "SEG-grounding": 0.0 } }, { "index": 1045, "category": "stamp_lora_gain", "instruction": "Please segment the object this sentence describes: \"red sox baseball player posing for picture with right hand on end of bat that is to his right side\".", "ious": { "Qwen-init": 0.8401373877657687, "STAMP-LoRA-init": 0.837453531598513, "SEG-grounding": 0.8371484389398798 } }, { "index": 808, "category": "seg_grounding_gain", "instruction": "Please segment the object this sentence describes: \"a asian woman wholding her hand to her face\".", "ious": { "Qwen-init": 0.0, "STAMP-LoRA-init": 0.0, "SEG-grounding": 0.0 } }, { "index": 3822, "category": "seg_grounding_gain", "instruction": "Please segment the object this sentence describes: \"a cow statue with two other cows standing on its back\".", "ious": { "Qwen-init": 0.005878735702865576, "STAMP-LoRA-init": 0.004945503433043646, "SEG-grounding": 0.018551152600495332 } }, { "index": 1641, "category": "seg_grounding_regression", "instruction": "Please segment the object this sentence describes: \"monitor closer to the mouse\".", "ious": { "Qwen-init": 0.0, "STAMP-LoRA-init": 0.0, "SEG-grounding": 0.0 } }, { "index": 2162, "category": "seg_grounding_regression", "instruction": "Please segment the object this sentence describes: \"there is a grey colored sofa chair in front of three people\".", "ious": { "Qwen-init": 0.7193003052359709, "STAMP-LoRA-init": 0.8150576082638061, "SEG-grounding": 0.6546229181652188 } }, { "index": 105, "category": "largest_method_disagreement", "instruction": "Please segment the object this sentence describes: \"back image of a woman in a green shirt\".", "ious": { "Qwen-init": 0.0010589775446730094, "STAMP-LoRA-init": 0.0006794162053346754, "SEG-grounding": 0.0006869800524450625 } }, { "index": 1021, "category": "largest_method_disagreement", "instruction": "Please segment the object this sentence describes: \"german shephard dog sleeping\".", "ious": { "Qwen-init": 0.8497363796133568, "STAMP-LoRA-init": 0.8622080375998349, "SEG-grounding": 0.8624919098959731 } }, { "index": 3778, "category": "all_methods_succeed", "instruction": "Please segment the object this sentence describes: \"the pizza that is closer to the camera than the other . it is in front of the glasses and the water\".", "ious": { "Qwen-init": 0.95012587968157, "STAMP-LoRA-init": 0.9503128588754253, "SEG-grounding": 0.935734395750332 } }, { "index": 754, "category": "all_methods_succeed", "instruction": "Please segment the object this sentence describes: \"chocolate in a white mug\".", "ious": { "Qwen-init": 0.8922919149643838, "STAMP-LoRA-init": 0.8976021463305572, "SEG-grounding": 0.8803074262121297 } } ], "legend": { "green": "true positive", "red": "false positive / over-segmentation", "blue": "false negative / under-segmentation" } }