Emergent 3D Instance Segmentation from Self-Supervised Point Transformers

• URL: http://arxiv.org/abs/2608.15796v1
• Date: Sun, 16 Aug 2026 15:20:57 GMT
• Title: Emergent 3D Instance Segmentation from Self-Supervised Point Transformers
• Abstract summary: Unsupervised 3D instance segmentation of outdoor LiDAR scans has traditionally relied on handcrafted geometric priors.<n>We investigate whether a frozen, self-supervised point transformer already contains the structural information required to isolate object instances.<n>We put these findings into our method TokenGraph3D, a training-free segmenter that groups points via connected components on a key-similarity graph.
• Score: 15.864488266763976
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Unsupervised 3D instance segmentation of outdoor LiDAR scans has traditionally relied on handcrafted geometric priors such as density-based clustering, motion cues, or projected 2D detections. In this work, we investigate whether a frozen, self-supervised point transformer already contains the structural information required to isolate object instances without any handcrafted geometric prior. Using this transformer purely as a feature extractor, we probe its internal representations across the SemanticKITTI, nuScenes, and Waymo Perception datasets. Our analysis yields four core insights: (1) the instance signal concentrates in the attention queries and keys rather than in the values or final output features; (2) output features semantically collapse, merging adjacent same-class objects that the queries and keys keep distinct; (3) this instance signal is bimodal in depth, strongest at the shallowest and deepest encoder stages; and (4) this signal is driven predominantly by the rotary position encoding (RoPE), whose removal collapses its advantage. We put these findings into our method TokenGraph3D, a training-free segmenter that groups points via connected components on a key-similarity graph, using neither density-based clustering nor proximity priors. Under identical prior-free conditions, we substantially outperform output-feature baselines, making the emergent 3D instance structure visible.

Related papers

• USR-Drive: Unified Driving Scene Representation via Joint Denoising of 3D Gaussians and Boxes [57.4583999658488]
  Spatial representation learning for autonomous driving aims to map raw visual signals into structured 3D scene representations.<n>Existing methods treat dynamic reconstruction and instance-level perception as separate tasks.<n>We propose USR-Drive, a unified conditional generative framework that recovers dense dynamic geometry and instance-level object layouts within a shared scene representation.
  arXiv  Detail & Related papers  (2026-08-19T15:29:06Z)
• Understanding Geometric Representations in Self-Supervised Vision Transformers via Subspace Intervention [7.887469434417159]
  We introduce a controlled subspace intervention framework to investigate how self-supervised Vision Transformers (ViTs) encode dense geometric information.<n>By connecting internal encoding mechanics with downstream performance, these findings provide a basis for effective feature selection and lightweight decoder design.
  arXiv  Detail & Related papers  (2026-07-02T10:18:02Z)
• FAST3DIS: Feed-forward Anchored Scene Transformer for 3D Instance Segmentation [15.271467111162714]
  FAST3DIS is an end-to-end approach that effectively bypasses post-hoc clustering.<n>We introduce a 3D-anchored, query-based Transformer architecture built upon a foundational depth backbone.<n> Experiments on complex indoor 3D datasets demonstrate that our approach achieves competitive segmentation accuracy.
  arXiv  Detail & Related papers  (2026-03-27T00:45:31Z)
• S-LAM3D: Segmentation-Guided Monocular 3D Object Detection via Feature Space Fusion [0.0]
  Monocular 3D Object Detection represents a challenging Computer Vision task due to the nature of the input used.<n>We introduce a decoupled strategy based on injecting precomputed segmentation information priors and fusing them directly into the feature space for guiding the detection.<n>The proposed method is evaluated on the KITTI 3D Object Detection Benchmark, outperforming the equivalent architecture that relies only on RGB image features for small objects in the scene.
  arXiv  Detail & Related papers  (2025-09-07T10:14:56Z)
• AutoInst: Automatic Instance-Based Segmentation of LiDAR 3D Scans [41.17467024268349]
  Making sense of 3D environments requires fine-grained scene understanding. We propose to predict instance segmentations for 3D scenes in an unsupervised way. Our approach attains 13.3% higher Average Precision and 9.1% higher F1 score compared to the best-performing baseline.
  arXiv  Detail & Related papers  (2024-03-24T22:53:16Z)
• KP-RED: Exploiting Semantic Keypoints for Joint 3D Shape Retrieval and Deformation [87.23575166061413]
  KP-RED is a unified KeyPoint-driven REtrieval and Deformation framework.<n>It takes object scans as input and jointly retrieves and deforms the most geometrically similar CAD models.
  arXiv  Detail & Related papers  (2024-03-15T08:44:56Z)
• OA-DET3D: Embedding Object Awareness as a General Plug-in for Multi-Camera 3D Object Detection [77.43427778037203]
  We introduce OA-DET3D, a plug-in module that improves 3D object detection.<n> OA-DET3D boosts the representation of objects by leveraging object-centric depth information and foreground pseudo points.<n>We conduct extensive experiments on the nuScenes dataset and Argoverse 2 dataset to validate the merits of OA-DET3D.
  arXiv  Detail & Related papers  (2023-01-13T06:02:31Z)
• SASA: Semantics-Augmented Set Abstraction for Point-based 3D Object Detection [78.90102636266276]
  We propose a novel set abstraction method named Semantics-Augmented Set Abstraction (SASA) Based on the estimated point-wise foreground scores, we then propose a semantics-guided point sampling algorithm to help retain more important foreground points during down-sampling. In practice, SASA shows to be effective in identifying valuable points related to foreground objects and improving feature learning for point-based 3D detection.
  arXiv  Detail & Related papers  (2022-01-06T08:54:47Z)
• M3DSSD: Monocular 3D Single Stage Object Detector [82.25793227026443]
  We propose a Monocular 3D Single Stage object Detector (M3DSSD) with feature alignment and asymmetric non-local attention. The proposed M3DSSD achieves significantly better performance than the monocular 3D object detection methods on the KITTI dataset.
  arXiv  Detail & Related papers  (2021-03-24T13:09:11Z)
• Monocular 3D Object Detection with Sequential Feature Association and Depth Hint Augmentation [12.55603878441083]
  FADNet is presented to address the task of monocular 3D object detection. A dedicated depth hint module is designed to generate row-wise features named as depth hints. The contributions of this work are validated by conducting experiments and ablation study on the KITTI benchmark.
  arXiv  Detail & Related papers  (2020-11-30T07:19:14Z)
• Unsupervised Object Detection with LiDAR Clues [70.73881791310495]
  We present the first practical method for unsupervised object detection with the aid of LiDAR clues. In our approach, candidate object segments based on 3D point clouds are firstly generated. Then, an iterative segment labeling process is conducted to assign segment labels and to train a segment labeling network. The labeling process is carefully designed so as to mitigate the issue of long-tailed and open-ended distribution.
  arXiv  Detail & Related papers  (2020-11-25T18:59:54Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.