Retrieving Objects from 3D Scenes with Box-Guided Open-Vocabulary Instance Segmentation

Conference: AAAI 2026 arXiv: 2512.19088 Code: https://github.com/ndkhanh360/BoxOVIS Area: 3D Vision Keywords: Open-vocabulary 3D instance segmentation, 3D object retrieval, superpoints, YOLO-World, 2D-to-3D lifting

TL;DR

This paper proposes a box-guided approach that leverages 2D bounding boxes from the open-vocabulary detector YOLO-World to guide the assembly of 3D instance masks from superpoints, without relying on SAM or CLIP. The method achieves high efficiency (<1 min/scene) while substantially improving retrieval performance on rare-category objects.

Background & Motivation

State of the Field

Open-vocabulary 3D instance segmentation (OV-3DIS) aims to retrieve objects of arbitrary categories from 3D point clouds given text queries, and is a core problem in robotics and augmented reality. Existing methods fall into two main categories:

• SAM+CLIP-based methods (OpenMask3D, Open3DIS, OVIR-3D): generate 2D masks with SAM → lift to 3D → classify with CLIP. Reasonable accuracy but extremely slow (5–10 min/scene), making deployment impractical.
• Efficient method Open-YOLO 3D: uses the class-agnostic 3D segmentor Mask3D to generate candidate masks and YOLO-World for classification, ~22 s/scene, eliminating SAM and CLIP.

Limitations of Prior Work

Despite its speed, Open-YOLO 3D has a critical weakness: it relies entirely on Mask3D (a pretrained 3D segmentor) to generate 3D candidate masks. Due to limited 3D training data (e.g., incomplete category coverage in ScanNet), Mask3D frequently misses low-frequency or rare categories (e.g., calendars, thermometers). Although YOLO-World can recognize such objects, Open-YOLO 3D uses it only for classification and not for generating new masks.

Root Cause & Starting Point

The tension lies between the limited generalization of 3D segmentors and the rich world knowledge of 2D detectors. The paper's core idea is to use 2D detection boxes to guide the assembly of new 3D instance masks from superpoints—inheriting the generalization capacity of 2D models while avoiding SAM and thus preserving efficiency.

Method

Overall Architecture

Input: 3D point cloud \(P\) + multi-view RGB-D images + camera intrinsics/extrinsics + text query. Output: 3D instance masks matching the query.

Pipeline: 1. Generate 3D superpoints (geometrically consistent regions) via graph-based segmentation. 2. Generate point-based masks using Mask3D (conventional path). 3. Run YOLO-World on RGB frames to produce 2D bounding boxes. 4. Box-Guided RGBD-Based Mask Generation: lift 2D boxes to 3D and assemble new instance masks from superpoints. 5. Merge both types of masks; use detection box results for classification.

Key Designs

1. Box-Guided RGBD-Based Mask Generation

Function: Generate 3D masks for rare objects missed by the 3D segmentor.

Core pipeline:

(a) Lifting 2D boxes to 3D: - For each RGB frame, YOLO-World produces detections \(B_i = \{(b_{ij}, c_{ij})\}\). - Pixels within each box are projected to 3D via depth maps and camera parameters. - Open3D is used to compute a 3D oriented bounding box \(b_{ij}^{3D}\) enclosing all projected points.

(b) Redundancy filtering: - If the overlap between a 3D box and any existing point-based mask exceeds \(\tau_{\text{box}}\%\), the object has already been detected by the 3D segmentor and the box is discarded.

(c) Superpoint assembly: - Superpoints are assigned to a box if \(\geq \tau_{\text{spp}}\%\) of their points fall inside the box. - This yields a coarse mask \(S_{ij}\) for each box.

(d) Cross-frame merging: - Frames are processed sequentially; if a new mask has IoU \(\geq \tau_{\text{merge}}\) with an existing candidate of the same category, their superpoints are merged; otherwise the new mask is added as a new candidate. - A final filtering pass discards new masks whose IoU with any point-based mask exceeds \(\tau_{\text{filter}}\), preserving the geometrically superior point-based masks for already-detected objects.

Design Motivation: - Superpoint assembly instead of SAM: superpoints are computed via the efficient Felzenszwalb graph-based algorithm, incurring far lower computational cost than SAM. - Redundancy filtering ensures the new masks complement rather than replace the 3D segmentor's output.

2. Box-Based Mask Classification

Function: Assign category labels to each 3D candidate mask.

Following Open-YOLO 3D, CLIP is not used: - Label map construction: for each frame, box regions are filled with detected category labels; larger boxes are filled first, then overwritten by smaller ones (intuition: a visible small object is closer to the camera than a larger one). - Visibility computation: all 3D points are projected to all frames at once to compute in-frame and occlusion-aware visibility. - Category aggregation: for each 3D mask, projected points in the top-\(k\) most visible frames are polled for their label map entries, and the most frequent category is assigned.

Loss & Training

This is a training-free / zero-shot method requiring no additional training. Pretrained components used: - Mask3D: class-agnostic 3D instance segmentor pretrained on ScanNet. - YOLO-World extra-large: open-vocabulary 2D detector. - Graph-based segmentation: the classical Felzenszwalb & Huttenlocher (2004) algorithm.

Inference settings: - ScanNet200: YOLO-World is applied to every 10th frame. - Replica: all frames are processed. - Images are downsampled by a factor of 5 during superpoint generation for efficiency.

Key Experimental Results

Main Results

ScanNet200 validation set:

Method	SAM	CLIP	mAP	mAP50	mAP25	mAP_tail	Time/scene
OpenMask3D	✓	✓	15.4	19.9	23.1	14.9	553.87s
Open3DIS	✓	✓	23.7	29.4	32.8	21.8	360.12s
Open-YOLO 3D	×	×	24.7	31.7	36.2	21.6	21.8s
Ours	×	×	24.9	32.1	36.8	22.4	55.9s

• Compared to Open-YOLO 3D: mAP +0.2, mAP50 +0.4, mAP25 +0.6, tail mAP +0.8.
• Speed (55.9s) is slower than Open-YOLO 3D (21.8s) but far faster than SAM/CLIP-based methods (360s+).

Replica dataset:

Method	mAP	mAP50	mAP25	Time/scene
OpenMask3D	13.1	18.4	24.2	547.32s
Open3DIS	18.5	24.5	28.2	187.97s
Open-YOLO 3D	23.7	28.6	34.8	16.6s
Ours	24.0	31.8	37.4	43.7s

Gains on Replica are more pronounced: mAP50 +3.2, mAP25 +2.6.

Ablation Study

No formal ablation table is presented; key observations are extracted from comparisons and discussion:

Configuration	Key Change	Effect
Point-based masks only (Open-YOLO 3D)	No RGBD-based masks	tail mAP 21.6; rare objects missed
+ Box-guided RGBD masks (Ours)	Novel instance discovery added	tail mAP 22.4 (+0.8); rare objects such as "calendar" detectable
RGBD mask quality	Superpoint assembly vs. SAM	Larger gains at IoU 50/25 thresholds; lower quality under strict IoU

Key Findings

1. Tail categories are the key gap: differences on head categories are minimal (slightly negative, −0.2), while tail categories show clear improvement (+0.8), confirming the core hypothesis that 3D segmentors generalize poorly to rare objects.
2. Gains are larger at lower IoU thresholds: improvement diminishes from mAP25 > mAP50 > mAP, since superpoint-assembled masks have coarser boundaries than SAM-based masks.
3. Qualitative results explicitly demonstrate that objects such as "calendar"—completely undetected by Open-YOLO 3D—are successfully retrieved by the proposed method.

Highlights & Insights

1. Clear design philosophy: rather than optimizing every component, the method strikes a practical balance between efficiency and generalization.
2. No additional training required: the entire pipeline is zero-shot, combining existing pretrained models without any training on 3D data.
3. Superpoints as a SAM substitute: an elegant solution to the 2D-to-3D mask lifting efficiency problem, relying on the classical and highly efficient Felzenszwalb graph-based segmentation.
4. Incremental design: new masks supplement rather than replace point-based masks, preserving the geometric accuracy of the 3D segmentor for common categories.

Limitations & Future Work

1. Speed bottleneck: primarily caused by Open3D's computation of 3D oriented bounding boxes; GPU-accelerated implementations are identified as an important direction.
2. Mask quality: superpoint-assembled masks are insufficiently precise under high IoU thresholds; future work could explore applying SAM refinement exclusively to final candidates.
3. Insufficient ablation: sensitivity analyses for hyperparameters (\(\tau_{\text{box}}, \tau_{\text{spp}}, \tau_{\text{merge}}, \tau_{\text{filter}}\)) are absent.
4. Indoor-only evaluation: both ScanNet200 and Replica are indoor datasets; generalization to outdoor scenes (e.g., autonomous driving) remains unvalidated.
5. The absolute mAP gain is modest (+0.2); the primary contribution lies in improving tail-category performance.

Related Work & Insights

• Open-YOLO 3D (ICLR 2025): the direct predecessor of this work, demonstrating the feasibility of eliminating SAM and CLIP.
• Open3DIS (CVPR 2024): pioneered the fusion of point-based and RGBD-based masks but relies on SAM.
• YOLO-World (CVPR 2024): real-time open-vocabulary detector serving as the core 2D module in this method.
• Felzenszwalb graph segmentation (2004): a 20-year-old classical algorithm finds renewed relevance in modern 3D understanding pipelines.
• Future work could explore replacing the 2D detector-based classifier with 3D open-vocabulary classifiers such as OpenShape or DuoMamba.

Rating

• Novelty: ⭐⭐⭐ — Straightforward yet effective; a component-level improvement rather than a paradigm shift.
• Experimental Thoroughness: ⭐⭐⭐ — Validated on two datasets but lacks ablations; performance gains are modest.
• Writing Quality: ⭐⭐⭐⭐ — Clear and concise, with well-motivated problem framing.
• Value: ⭐⭐⭐⭐ — Strong practical utility; provides an efficient solution for retrieving rare objects in 3D scenes.

Related Papers

• [AAAI 2026] OpenScan: A Benchmark for Generalized Open-Vocabulary 3D Scene Understanding
• [AAAI 2026] UniC-Lift: Unified 3D Instance Segmentation via Contrastive Learning
• [AAAI 2026] ASSIST-3D: Adapted Scene Synthesis for Class-Agnostic 3D Instance Segmentation
• [ICLR 2026] GeoPurify: A Data-Efficient Geometric Distillation Framework for Open-Vocabulary 3D Segmentation
• [CVPR 2026] JOPP-3D: Joint Open Vocabulary Semantic Segmentation on Point Clouds and Panoramas