adu2021 Component ID
Group-Gated Fragment Merging with Semantic Compatibility Constraints
Motivation
Open-vocabulary 3D detection from multi-view RGB images typically decouples instance construction from semantic labeling. Fragments are merged based on geometric consistency first, then labeled. This causes irreversible errors when geometric evidence is incomplete or view-dependent, and the approach cannot recover from geometry-driven over-merging decisions.
The Modification
Group3D integrates semantic constraints into the merge decision itself, requiring both semantic compatibility and geometric consistency before combining fragments:
Scene-Adaptive Vocabulary Memory - Query a multimodal large language model (MLLM) across views to identify object categories present in the scene, then aggregate into a scene-specific vocabulary.
Semantic Compatibility Groups - The MLLM partitions the vocabulary into groups capturing "plausible cross-view category equivalence" under taxonomy noise. Groups identify categories that could refer to the same physical object while excluding structurally incompatible associations.
Group-Gated Fragment Merging - Two 3D fragments merge only when they satisfy both conditions:
- Semantic gate: Categories belong to the same compatibility group
- Geometric gate: Voxel-level spatial overlap meets thresholds
The algorithm processes fragments sorted by spatial extent, attempting merges with existing clusters while respecting both gates.
Ablation Results
The paper demonstrates:
- Semantic grouping alone improves over geometry-only merging by preventing false positives on geometrically overlapping but semantically distinct objects
- Multi-view evidence accumulation (cross-view confidence aggregation) provides consistent improvements
- The approach prevents catastrophic geometric over-merging while maintaining high recall
- Performance gains are consistent across different scene complexities and view counts
Conditions
- Requires multimodal LLM access for vocabulary and compatibility group generation
- Works best when geometric evidence is incomplete or ambiguous (where semantic constraints have highest impact)
- Assumes sufficient multi-view coverage to accumulate meaningful cross-view evidence
- Scene-adaptive group generation requires per-scene inference time (amortized cost)
Drop-In Checklist
- Integrate MLLM query module to extract scene vocabulary from multi-view images
- Implement semantic compatibility group generation (MLLM partitions vocabulary)
- Add group membership check to fragment merge decision logic
- Implement group-gated threshold for geometric consistency checks
- Add multi-view evidence aggregation for final instance labeling
- Validate that semantic gates prevent geometry-driven over-merging
- Benchmark detection accuracy (precision/recall) on open-vocabulary 3D detection benchmarks