Frame-to-Scan Matching on ScanNet-SG-GPT Difficult (test)
42AccuracyOurs H DGSA + MCF Final
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Ours H DGSA + MCF FinalModel Scale=High-performance, Encoder Architecture=DGSA, Matching Allocator=MCF Final2026.05 | 42 | 44 | 67.1 | 50.3 | |
| Ours H DGSA + MNNModel Scale=High-performance, Encoder Architecture=DGSA, Matching Allocator=MNN2026.05 | 41.6 | 44 | 66.1 | 50 | |
| Ours H FFN + MNNModel Scale=High-performance, Encoder Architecture=FFN, Matching Allocator=MNN2026.05 | 39.4 | 42.1 | 62.8 | 47.6 | |
| Ours H T-GNN + MNNModel Scale=High-performance, Encoder Architecture=T-GNN, Matching Allocator=MNN2026.05 | 39.1 | 41.2 | 58.7 | 45.8 | |
| VLM+Bert CosSimFeature Embedding Type=VLM+Bert2026.05 | 35.6 | 36.3 | 74 | 46.2 | |
| Ours L DGSA + MCF FinalModel Scale=Lightweight, Encoder Architecture=DGSA, Matching Allocator=MCF Final2026.05 | 35.6 | 35.6 | 78.8 | 46.6 | |
| Ours L DGSA + MNNModel Scale=Lightweight, Encoder Architecture=DGSA, Matching Allocator=MNN2026.05 | 35.5 | 37.2 | 67.6 | 42.2 | |
| VLM CosSimFeature Embedding Type=VLM2026.05 | 30.6 | 35.4 | 50.4 | 38.1 | |
| Bert CosSimFeature Embedding Type=Bert2026.05 | 26 | 33.9 | 37.7 | 30.6 | |
| Ours L T-GNN + MNNModel Scale=Lightweight, Encoder Architecture=T-GNN, Matching Allocator=MNN2026.05 | 24.2 | 25.4 | 52.9 | 23.2 | |
| ROMAN2026.05 | 14.3 | 18.7 | 15.6 | 16 | |
| SG-Reg2026.05 | 1.2 | 1.9 | 3.3 | 2.4 |