An arXiv preprint reports higher benchmark scores for a system designed to detect 3D objects from categories not treated as base classes. Called Co-3DGT, the method recorded novel-class average precision, or AP, of 14.37% on SUN RGB-D and 21.91% on ScanNetV2. The paper reports those figures as 4.71 and 9.82 percentage points higher, respectively, than OV-Uni3DETR.
Co-3DGT addresses open-vocabulary 3D object detection, a task in which the benchmark separates object categories into base and novel classes. The paper says its goal is to make novel-object discovery more reliable and detector training more robust. In practical terms, the system is being tested not only on categories represented in the base split, but also on novel categories.
The framework is organized into a discovery stage and a training stage. Discovery uses co-distillation, while training combines uncertainty regularization with hierarchical semantic alignment. The approach therefore brings together a way to identify novel objects and a training strategy intended to make the detector more robust.
The benchmark signal
Co-distillation links 2D bounding boxes with 3D proposals by treating the associations as a bipartite matching problem, then solving that problem with the Hungarian algorithm. This matching step is the bridge between the two forms of scene information used in the discovery stage.
The tests used SUN RGB-D and ScanNetV2. SUN RGB-D uses 5,285 training samples in the reported setup, with 10 base classes and 36 novel classes. ScanNetV2 uses 1,201 training scenes, with 10 base classes and 50 novel classes. These splits allow the paper to report separate performance for base categories, novel categories and the combined set.
The paper evaluates mean average precision and mean average recall at an intersection-over-union, or IoU, threshold of 0.25. IoU refers to the overlap between a predicted object box and the reference box; the threshold sets the overlap requirement used for the reported scores.
On the novel-class AP measure, Co-3DGT reached 14.37% on SUN RGB-D and 21.91% on ScanNetV2. Against OV-Uni3DETR, the reported gains were 4.71 percentage points and 9.82 percentage points. The result was reported on both datasets.
Across the wider class set, the paper reports APmean of 22.63% on SUN RGB-D and 23.75% on ScanNetV2. It gives gains of 4.57 and 8.60 percentage points over OV-Uni3DETR. The reported difference therefore was not limited to the novel-class score; the overall mean also favored Co-3DGT in the two benchmark comparisons.
The paper adds a separate annotation-free check on ScanNetV2. The evaluation uses no ground-truth annotations and covers the 20 most frequent unseen classes. In that setting, Co-3DGT’s APmean was 36.15%, with a reported improvement of 10.82 percentage points over OV-Uni3DETR. Because this test uses a different setup, its number should be read alongside, rather than substituted for, the main class-split results.
What the ablations add
An ablation examined the discovery stage on its own. Compared with the 3D-Proposal-based pipeline, Co-distil(CuTR) reported a 2.99-point gain in precision and a 10.72-point gain in recall on SUN RGB-D. On ScanNetV2, the reported gains were 5.89 points in precision and 12.01 points in recall.
Another comparison shows what the discovery variant looked like before the robust-training methods were included. Co-distil(CuTR) produced novel-class AP of 12.69% on SUN RGB-D and 18.74% on ScanNetV2 in that configuration. Those figures are below the novel-class AP values reported for Co-3DGT in the main comparison, although the reviewed analysis gives no uncertainty intervals or significance tests for the difference.
The training-side ablation focused on the combined uncertainty-regularization and hierarchical-alignment variant. It reported APmean of 22.63% on SUN RGB-D and 23.75% on ScanNetV2, with gains of 2.01 and 3.60 percentage points. Alongside the discovery ablation, this comparison shows how the paper assessed the framework’s proposed parts as well as the combined system.
What the results do not settle
The narrow conclusion is that Co-3DGT reported higher novel-class and overall benchmark scores than OV-Uni3DETR on SUN RGB-D and ScanNetV2, plus a higher score in the separate annotation-free test. The stronger claim that the method will generalize to additional datasets or scene types has not been established by the reported experiments.
Open questions remain about real-time operation with temporal fusion and about performance under alternative detector and language-model choices. The reviewed evidence also reports no confidence intervals or significance tests, so the numerical gaps remain reported benchmark differences rather than statistical proof of superiority.
The document is an arXiv version 1 preprint posted on 20 August 2026. It reports support from the National Key Research and Development Program of China and the Ministry of Education, Singapore, under the listed grants. For now, its contribution is best understood as a method proposal evaluated through benchmark comparisons.
Paper data and sources
Original title: Open-Vocabulary 3D Object Detection with Co-Distillation Discovery and Dual Guidance Robust Training
Authors: Shangbo Yuan, Jie Xu, Xiaofeng Zhu, Na Zhao
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text