Source-linked AI summary
TopoField: Technical Report for ECCV 2026 SceneFun3D Challenge
Zhiwei Chen, Zhiheng Fu, Zixu Li, Mingyu Zhang, Jiale Huang, Yupeng Hu
TL;DR
Small, partly occluded functional parts must be recovered as confidence-ranked masks, although point-wise predictors produce continuous probability fields. TopoField converts test-time predictions into a semantic consensus field and decodes and ranks instance candidates; it reports 39.67% validation AP and 28.286% official-test AP.
Problem
Small, partly occluded functional parts make point-wise probabilities noisy, while the benchmark evaluates ranked discrete masks rather than continuous fields.
Source: §1 · p. 2Method
Results
39.67% validation AP and 28.286% official-test AP are reported for TopoField.
Source: §5 · p. 9Takeaways & Limitations
Validation analysis treats candidate-pool coverage and top-ranked mask quality as complementary measures.
Source: §1 · p. 2
Abstract
from arXiv · showhide
Functional segmentation requires locating small interactive parts, such as knobs, buttons, and sockets, in indoor point clouds and returning confidence-ranked instance masks. A point-wise semantic predictor produces a continuous probability field, whereas the benchmark evaluates a discrete set of scored instances. This creates a substantial field-to-instance gap. We present TopoField, a deterministic decoder built on top of VOLT. TopoField first converts 13 test-time forward passes into a semantic consensus field, then decodes instance candidates from the consensus geometry, and finally ranks the candidates in a class-aware manner. The augmentation stage retains the processed-point identity of every fragment prediction, accumulates probabilities in a common point array, normalizes by coverage count, and fuses the resulting caches. The decoding stage separates high-confidence cores from broader support regions, constructs a symmetric radius-pruned point graph, and uses a multi-source graph-geodesic partition to split support regions that contain multiple seeds. On 561 ground-truth instances across 30 validation scenes, an IoU-0.5-valid candidate exists for 67.2% of instances, and top-1 coverage is 20.0%. These complementary measurements motivate joint attention to candidate construction and ranking. TopoField achieves 28.286% AP, 41.383% AP50, and 47.558% AP25. TopoField achieved second place on the ECCV 2026 SceneFun3D Challenge.
1 Introduction
Functional parts are small and often occluded, while point-wise predictors produce probability fields rather than the ranked masks the benchmark evaluates. TopoField addresses this gap with index-aligned semantic consensus and geometry-aware instance decoding.
1 Introduction
- Small, partly occluded parts make point probabilities noisy, and proposal construction and ordering can create field-to-instance error despite reasonable semantics.High thresholds can truncate weak boundaries, whereas low thresholds can merge targets with hosts or neighbors.Source: §1 · p. 2
- An IoU-0.5-valid candidate exists for 67.2% of 561 instances, while top-1 coverage is 20.0% across 30 validation scenes.The statistics characterize candidate-pool coverage and overlap from top-ranked candidates.Source: §1 · p. 2
- TopoField combines index-aligned test-time semantic consensus with geometry-aware decoding that selects class-specific operators and scores candidates.The consensus stage aggregates and normalizes fragment predictions; the seeded graph-basin branch partitions support points by graph-path length.Source: §1 · p. 2
- The paper frames its contribution around the gap between point-wise semantics and a ranked instance-set output, quantified with complementary coverage statistics.Source: §1 · p. 2
- 28.286% AP and 41.383% AP50 are reported for the parameter-free decoder.Source: §1 · p. 2
2 Related Work
Related work spans learned 3D instance grouping, functional-part segmentation, graph-based partitioning, and test-time augmentation. TopoField instead uses a fixed point-wise predictor and deterministic grouping to construct and rank instance proposals.
2 Related Work
- PointGroup and SoftGroup group point-wise semantics, while Mask3D and SPFormer use learned instance or superpoint queries to predict masks.These approaches jointly learn representation, grouping, and refinement.Source: §2 · p. 3
- SceneFun3D extends indoor scene understanding to small, often occluded functional parts; VOLT supplies TopoField’s point-wise semantics.Source: §2 · p. 3
- Graph-based region methods propagate labels from seeds over local adjacency, and TopoField uses cores and radius-pruned graph paths to partition support regions.Connected components isolate disconnected supports before geodesic propagation assigns points in connected supports with several cores.Source: §2 · p. 3
- Test-time augmentation aggregates predictions from transformed inputs, but per-point prediction requires preserving point identity across augmentation and voxelization.TopoField scatters probabilities through retained indices, normalizes by coverage, and fuses caches before decoding.Source: §2 · p. 3
- Prior noisy-correspondence work emphasizes evidence selection and confidence estimation, which TopoField applies through semantic consensus, core selection, and class-aware scoring.Source: §2 · p. 4
3 Method
TopoField is a deterministic decoder that turns point-wise predictions into foreground instance masks and scores. It stabilizes predictions across test-time passes, then uses class-conditioned geometry-aware decoding to form and rank candidates.
3.1 Problem formulation
- TopoField emits foreground instance masks in original point order, each paired with a target class and a score for ordering candidates.The predictor assigns probabilities across background and nine target categories.Source: §3.1 · pp. 4–5
3.2 Stage 1: TTA-based Semantic Consensus
- Across 13 passes, retained processed-point indices align predictions across coordinate and normal transformations; softmax probabilities accumulate in a common N×10 array, are averaged by contribution count, and independent caches are fused.All 13 passes receive equal weights; indices 12 and 2 duplicate geometrically, giving the duplicated outcome a combined weight of 2/13.Source: §3.2 · p. 5
3.3 Stage 2: Geometry-aware Instance Decoding
- For each category, the decoder selects among argmax connectivity, foreground-threshold connectivity, and a seeded graph basin.The basin route separates high-confidence core points from broader support regions, using thresholds τs = 0.15 and τh = 0.40 in the representative recipe.Source: §3.3 · pp. 5–6
- The seeded basin builds a symmetric, radius-pruned graph over support points and assigns points in multi-seed components by graph-geodesic distance.The implementation uses k = 32 neighbors and edge weights based on Euclidean distance, bounded below by 10^-12.Source: §3.3 · p. 6
- Candidates are ranked primarily by mean consensus probability for their class, with optional quantile or hybrid pooling and class-wise size scaling, then same-class overlaps are suppressed.When αc = 0, the size-scaled score reduces to the unscaled score; exact duplicate masks retain the highest-scoring copy.Source: §3.3 · pp. 6–7
3.4 Implementation details
- VOLT supplies the point-wise predictions, while inference uses frozen probability caches without gradient updates; Stage 1 is fixed when Stage 2 is varied.The implementation uses 1 cm voxels and input features XYZ, RGB, and normals.Source: §3.4 · p. 7
4 Experiments
Experiments assess candidate coverage and ranking separately, compare decoder configurations, and report validation and official-test performance. Support/core selection, graph-geodesic partitioning, and class-conditioned decoding contribute to the reported validation results.
4.1 Setup
- SceneFun3D evaluation uses 160 training, 30 validation, and 30 test scenes across nine categories, with AP over IoU 0.50–0.95 as the official metric.The setup also reports AP50 and AP25.Source: §4.1 · p. 7
4.2 Candidate diagnosis
- Using fixed epoch-80 EMA 13-TTA caches, Table 1 reports a 47.2-point gap between any-candidate and top-1 coverage, distinguishing candidate-pool coverage from top-ranked mask quality for Stage 2 evaluation.The measurements cover 561 ground-truth instances in 30 validation scenes.Source: §4.2 · p. 7, Table 1 · p. 7
4.3 Validation and test results
- 28.286% AP, 41.383% AP50, and 47.558% AP25 are achieved by the epoch-100 aggressive configuration on the official test.Source: §4.3 · p. 8, Table 3 · p. 8
4.4 Discussion
- Full-validation AP rises from 36.58 to 39.67 across decoder components, with support/core selection contributing the largest development-split gain.Graph-geodesic partitioning adds 0.71 points and class-conditioned decoding adds 0.85 points on the development split.Source: §4.4 · p. 8, §4.3 · p. 8, Table 2 · p. 8
- Support/core selection retains boundary evidence and supplies confident seeds; graph-geodesic assignment partitions multi-core support regions by 3D connectivity.The two stages aggregate index-aligned point predictions into a semantic consensus field, then construct and rank instance masks.Source: §4.4 · p. 8
5 Conclusion
TopoField is a two-stage deterministic decoder for functional interaction element segmentation. It builds a semantic field from multiple forward passes, decodes and ranks instance masks, and achieves 39.67% validation AP and 28.286% official-test AP.
- TopoField combines index-aligned semantic consensus with class-conditioned instance decoding and graph-geodesic assignment of support points to seed components.The decoder adds no learnable parameters to the point-wise predictor.Source: §5 · pp. 8–9
- The combined decoder achieves 39.67% validation AP and 28.286% official-test AP.Source: §5 · p. 9