COGS: A Diagnostic Benchmark for Attribute-Aware Object-Level 3D Gaussian Splatting Perception

Zheng Xiao; Yifan Liu; Weijian Deng; Jiashu Li; Ziqin Liu; Zhenjun Han; Jianbin Jiao

cs.CVOct 29, 2026

COGS: A Diagnostic Benchmark for Attribute-Aware Object-Level 3D Gaussian Splatting Perception

Authors: Zheng Xiao, Yifan Liu, Weijian Deng, Jiashu Li, Ziqin Liu, Zhenjun Han, Jianbin Jiao

Organizations: School of Advanced Interdisciplinary Sciences, University of Chinese Academy of Sciences, China · University of Chinese Academy of Sciences, China · Tsinghua Shenzhen International Graduate School, China

Abstract

3D Gaussian Splatting (3DGS) is increasingly explored as an input representation for 3D perception. Yet it remains unclear how current 3D architectures exploit heterogeneous Gaussian attributes. This ambiguity makes accuracy hard to interpret and limits principled architecture design for 3DGS perception. However, existing object-level 3DGS datasets lack the appearance variation or artifact control needed for systematic attribute-level analysis. We introduce COGS, a large-scale 3DGS dataset with 15,675 artifact-filtered and appearance-rich objects across 40 categories. Built on COGS, we establish COGS-Bench for object classification and attribute-usage diagnosis, complemented by external validation. Our benchmark shows that Gaussian attributes are not uniformly beneficial: selected subsets can outperform the full attribute set, and the effects of appearance, opacity, and covariance-related attribute combinations vary substantially across architectures. We further study this finding with HALO, a reference model that preserves attribute-specific evidence before hierarchical aggregation. Experiments on COGS-Bench and external 3DGS datasets demonstrate that COGS supports systematic diagnosis of 3DGS attribute usage, while HALO highlights the value of role-preserving attribute organization. Our dataset is available at https://huggingface.co/datasets/XiaoZhengCV/COGS.

Figures & tables

Figure 1: Overview of COGS and COGS-Bench. Existing object-level 3DGS datasets provide complementary but limited properties, motivating COGS as a curated, artifact-filtered, and appearance-rich dataset for controlled attribute diagnosis.

Figure 2: Overview of the COGS construction pipeline. Starting from large-scale web image collection, we perform semantic localization, foreground-aware filtering and deduplication, manual image inspection, generative 3DGS reconstruction with TRELLIS, and viewer-based 3DGS quality inspection to obtain curated object-centric 3DGS assets.

Figure 3: Discarded 3DGS samples rendered from multiple views. Row 1 shows view-inconsistent structural duplication, where drawer features expected on one face also appear on adjacent faces; row 2 shows unintended fusion between two lipstick bodies; row 3 shows incomplete chair geometry.

Figure 4: Category distribution comparison. Categories in each dataset are independently sorted by object count. The x-axis represents dataset-specific sorted category indices rather than aligned category identities. MACGS statistics use the authors’ publicly released 30-category dataset.

Figure 5: Overview of HALO. Given a set of Gaussian primitives, HALO progressively builds hierarchical representations through repeated Feature Extraction (FE), Hierarchical Aggregation (HA), and downsampling. FE performs multi-stream attribute encoding with Locally-decOupled MLP (LOM), while HA aggregates neighborhood evidence before primitive decimation.

Dataset	48 views		Single view
Dataset	MUSIQ ( $\uparrow$ )	CLIPIQA+ ( $\uparrow$ )	MUSIQ ( $\uparrow$ )	CLIPIQA+ ( $\uparrow$ )
ModelNet40GS [ 18 ]	$58.71\,(\pm 3.39)$	$0.5228\,(\pm 0.0158)$	$52.25\,(\pm 5.33)$	$0.4969\,(\pm 0.0162)$
MACGS [ 37 ]	$38.64\,(\pm 3.87)$	$0.4509\,(\pm 0.0256)$	$49.94\,(\pm 5.23)$	$0.5172\,(\pm 0.0233)$
COGS	$\mathbf{64.75\,(\pm 3.84)}$	$\mathbf{0.6013\,(\pm 0.0278)}$	$\mathbf{63.21\,(\pm 4.54)}$	$\mathbf{0.5863\,(\pm 0.0343)}$

Table 1: Rendered visual quality comparison. We report MUSIQ and CLIPIQA+ scores on rendered 48-view and single-view images. Standard deviations are shown in parentheses.

Figure 6: Gaussian attribute statistics across COGS, MACGS, and ModelNet40GS. We compare primitive count, high-opacity ratio (

>

0.8), normalized Gaussian volume, and color saturation.

Figure 7: Intra-class pairwise cosine distance distribution in the DINOv2 feature space. COGS shows the largest spread, indicating richer intra-class diversity than ModelNet40GS and MACGS.

Method	OA (% $\uparrow$ )	mAcc (% $\uparrow$ )	Mem. (MB $\downarrow$ )	Time (min $\downarrow$ )	Params (M $\downarrow$ )
PointNet [ 20 ]	79.53	79.70	1112	17.40	3.47
DGCNN [ 29 ]	85.69	86.70	5411	61.09	1.81
PTv1 [ 38 ]	87.67	88.32	744	72.80	3.26
DeLA [ 5 ]	88.48	89.14	1771	33.06	5.34
PTv3 [ 30 ]	89.11	89.77	5710	212.15	38.87
Gaussian-MAE [ 18 ]	88.20	88.58	2876	49.31	22.10

Table 2: Classification benchmark on COGS with all Gaussian attributes as input.

Method	P	PO	PC	POC	PSQ	POSQ	PSQC	POSQC
PointNet [ 20 ]	73.25/68.48	80.73/79.84	76.52/75.82	77.27/77.24	81.23/79.31	82.71/81.30	78.93/78.49	79.53/79.70
DGCNN [ 29 ]	85.87/86.15	87.04/87.31	82.79/83.84	84.17/85.20	88.57/88.42	89.18/89.36	85.29/86.08	85.69/86.70
PTv1 [ 38 ]	86.45/86.43	87.10/87.65	87.85/88.34	88.09/88.76	88.54/88.27	88.82/89.26	88.46/89.15	87.67/88.32
DeLA [ 5 ]	89.28/89.97	91.21/91.68	83.17/84.31	85.45/86.54	91.17/91.33	90.40/91.00	87.58/88.49	88.48/89.14
PTv3 [ 30 ]	88.53/88.80	89.50/89.89	89.16/89.16	88.58/88.89	88.74/89.01	89.68/89.82	88.44/88.68	89.11/89.77
Gaussian-MAE [ 18 ]	87.01/86.32	89.18/89.49	85.89/85.79	86.56/87.43	91.71/91.09	89.10/88.96	87.01/87.18	88.20/88.58

Table 3: Attribute usage diagnosis on COGS-Bench. P, O, S, Q, and C denote position, opacity, scale, rotation, and appearance/color, respectively. Each entry reports OA/mAcc (%).

HALO-B	OA (% $\uparrow$ )	mAcc (% $\uparrow$ )
Shared attribute processing	89.03	89.67
Attribute-specific processing	94.50	94.70

Table 4: Ablation on attribute-specific processing in HALO-B.

Models	Diagnostic result (OA, %)
6 conventional baselines	6/6: subset $>$ POSQC (+0.28–3.05)
HALO-B / HALO-L	POSQC best: 93.99 / 94.55

Table 5: Cross-pipeline attribute diagnosis on MACGS (OA).

Method	Variant/Input	mIoU (%)	OA (%)
PTv1	POSQC	48.45	80.95
	POSQ	50.37	81.08
HALO	Shared POSQC	43.81	79.16
	Attribute-specific POSQC	49.83	81.96

Table 6: Scene-level validation on ScanNet.

Method	10%	20%	50%	100%
PointNet [ 20 ]	56.65 (+26.98)	64.33 (+18.58)	76.25 (+7.03)	81.89 (+2.13)
DGCNN [ 29 ]	60.54 (+21.91)	71.35 (+12.20)	81.79 (+6.00)	88.82 (+1.67)
PTv1 [ 38 ]	62.29 (+12.01)	73.29 (+8.41)	84.94 (+4.81)	90.11 (+1.85)
HALO-B	69.87 (+3.23)	82.72 (+2.04)	90.29 (+1.10)	94.08 (+0.09)

Table 7: Models are fine-tuned on varying fractions of MACGS using either random initialization or COGS pre-training. Each cell reports pre-trained OA and OA gain over training from scratch. Full results are provided in the supplementary material.

Figure S1 : Category distribution of the COGS dataset. The figure reports the 40 category names and the corresponding number of curated 3DGS assets in each category.

Method	Fraction	Scratch	Pre-trained on COGS
Method	Fraction	OA / mAcc (% $\uparrow$ )	OA / mAcc (% $\uparrow$ )
PointNet [ 20 ]	10%	29.67 / 29.12	56.65 (+26.98) / 54.51 (+25.39)
DGCNN [ 29 ]		38.63 / 37.28	60.54 (+21.91) / 58.42 (+21.14)
PTv1 [ 38 ]		50.28 / 49.73	62.29 (+12.01) / 60.99 (+11.26)
HALO-B (Ours, Ref.)		66.64 / 64.95	69.87 (+3.23) / 68.53 (+3.58)
PointNet [ 20 ]	20%	45.75 / 45.01	64.33 (+18.58) / 62.84 (+17.83)

Table S1 : Full cross-dataset transfer results from COGS to MACGS. Models are fine-tuned on varying fractions of MACGS using either random initialization or COGS pre-training. Gains over Scratch are shown in parentheses.

Method	P	PO	PC	POC	PSQ	POSQ	PSQC	POSQC
PointNet [ 20 ]	76.71/75.56	78.10/77.59	78.65/77.79	81.51 /80.83	77.36/76.24	76.06/75.65	78.16/77.61	79.76/77.97
DGCNN [ 29 ]	88.91/88.24	87.89/86.71	88.17/87.28	90.20 /89.06	87.62/86.31	86.04/84.87	87.71/87.08	87.15/85.80
PTv1 [ 38 ]	88.91/87.79	87.80/86.68	87.99/87.33	88.82/88.46	89.00/88.11	89.37/88.85	90.93 /89.53	88.26/87.54
DeLA [ 5 ]	91.50/90.43	91.50/90.47	92.51 /91.97	92.14/90.90	92.05/90.88	91.87/90.43	92.24/91.28	91.50/90.15
PTv3 [ 30 ]	89.93/89.31	90.76/89.72	92.51/91.74	92.61/91.82	92.51/91.72	92.79 /92.17	92.61/92.03	92.33/91.45
Gaussian-MAE [ 18 ]	87.62/86.59	89.37/88.76	90.30/89.08	91.50 /90.61	90.11/89.61	89.74/88.71	91.40/90.45	91.22/90.49

Table S2 : Complete attribute-usage diagnosis on MACGS. P, O, S, Q, and C denote position, opacity, scale, rotation, and appearance/color, respectively. Each entry reports OA/mAcc (%). Bold numbers indicate the best OA for each method.

Method	Variant/Input	mIoU (% $\uparrow$ )	OA (% $\uparrow$ )
PTv1 [ 38 ]	P	45.64	78.84
PTv1 [ 38 ]	PO	49.67	81.14
PTv1 [ 38 ]	PC	47.81	80.50
PTv1 [ 38 ]	POC	46.70	80.49
PTv1 [ 38 ]	PSQ	48.05	79.71
PTv1 [ 38 ]	POSQ	50.37	81.08

Table S3 : Real-world scene-level proof-of-concept on the standard ScanNet 20-class semantic segmentation task. Training uses the official ScanNet Data-Efficient 10% split and evaluation uses the complete validation set.

Method	OA (% $\uparrow$ )	Params (M $\downarrow$ )	Counted FLOPs (G $\downarrow$ )	B1 / B32 Lat. (ms/batch $\downarrow$ )	B32 Throughput (obj/s $\uparrow$ )	B1 / B32 Peak Mem. (MiB $\downarrow$ )
PointNet [ 20 ]	79.53	3.47	0.45	1.24 / 3.81	8394	34.24 / 303.31
DGCNN [ 29 ]	85.69	1.81	2.73	1.38 / 26.48	1208	76.11 / 1968.81
PTv1 [ 38 ]	87.67	3.26	0.10	10.48 / 18.35	1744	27.48 / 237.56
DeLA [ 5 ]	88.48	5.34	0.73	3.49 / 7.92	4039	46.33 / 565.85
PTv3 [ 30 ]	89.11	38.87	6.32	30.95 / 71.62	447	271.25 / 424.38
Gaussian-MAE [ 18 ]	88.20	22.10	2.44	4.01 / 8.84	3619	104.68 / 484.68

Table S4 : Inference efficiency under the unified 1,024-Gaussian classification protocol. Counted FLOPs are obtained using fvcore with one fused multiply-add counted as one FLOP; unsupported point-cloud-specific operators are excluded from FLOPs but included in latency and memory measurements. B1/B32 latency denotes the mean CUDA-event execution time of the complete batch after 100 warm-up and 1,000 measured forward passes. B32 throughput is computed from the measured B32 batch latency. Peak inference memory reports the maximum allocated GPU memory during one forward pass.

Cross-level Fusion	Neighborhood Agg.	OA (% $\uparrow$ )	mAcc (% $\uparrow$ )	Mem. (MB $\downarrow$ )
$\times$	–	84.26	85.15	482
✓	$\times$	85.07	85.68	603
✓	$\circ$	88.97	90.03	1885
✓	✓	94.50	94.70	750

Table S5 : Ablation study of hierarchical aggregation on COGS-Bench. “–” indicates that the component is not applicable.

\circ

denotes that downsampling is omitted.

Method	MACGS		ModelNet40GS
Method	OA (% $\uparrow$ )	mAcc (% $\uparrow$ )	OA (% $\uparrow$ )	mAcc (% $\uparrow$ )
PointNet [ 20 ]	79.76	77.97	90.98	87.79
DGCNN [ 29 ]	87.15	85.80	90.82	87.50
PTv1 [ 38 ]	88.26	87.54	91.06	88.17
DeLA [ 5 ]	91.50	90.15	91.99	89.60
PTv3 [ 30 ]	92.33	91.45	91.97	89.14

Table S6 : Full-POSQC cross-pipeline classification results on MACGS and ModelNet40GS.

Figure S2 : Additional multi-view visualizations of COGS samples.

Figure S3 : Additional multi-view visualizations of COGS samples.

Figure S4 : Additional multi-view visualizations of COGS samples.

Figure S5 : Additional multi-view visualizations of COGS samples.

Figure S6 : Additional multi-view visualizations of COGS samples.

Figure S7 : Additional multi-view visualizations of COGS samples.

3D Gaussian Splatting

3D Gaussian Splatting represents scenes as optimizable 3D Gaussian primitives for fast, photorealistic rendering from new viewpoints.

See all 1,027 papers

OP2GS: Object-Aware 3D Gaussian Splatting with Dual-Opacity Primitives
May 19, 2026Guiyu Liu, Niklas Vaara, Janne Mustaniemi +23D Gaussian Splatting Open-Vocabulary 3D Segmentation
Center for Machine Vision and Signal Analysis, University of Oulu, Finland · Aalto University, Finland
Learning Representations from 3D Gaussian Splats
May 28, 2026Julia Farganus, Krzysztof Żurawicki, Arkadiusz Gaweł +23D Gaussian Splatting Geometric Representation Learning
Department of Artificial Intelligence, Wrocław University of Science and Technology, Wrocław, Poland
Drop-In Perceptual Optimization for 3D Gaussian Splatting
Mar 23, 2026Ezgi Ozyilkan, Zhiqi Chen, Oren Rippel +23D Gaussian Splatting Perceptual Loss
Apple · New York University Tandon School of Engineering

May 19, 2026J/K move · Enter open · S save

OP2GS: Object-Aware 3D Gaussian Splatting with Dual-Opacity Primitives

Guiyu Liu, Niklas Vaara, Janne Mustaniemi, Juho Kannala+1

Center for Machine Vision and Signal Analysis, University of Oulu, Finland · Aalto University, Finland

3D Gaussian Splatting Open-Vocabulary 3D Segmentation 3D Scene Representation 3D Instance Segmentation 3DGS Segmentation

Open paper AskPDF arXiv

3D Gaussian Splatting (3DGS) provides an explicit and efficient scene representation, but its primitives lack inherent object-level identity, hindering downstream tasks such as open-vocabulary scene understanding. Existing methods typically address this by either distilling high-dimensional feature embeddings into Gaussians or by lifting 2D mask labels into 3D via heuristic refinement. However, feature-based approaches incur heavy storage and decoding overhead, while lifting-based pipelines remain vulnerable to label contamination: Gaussians necessary for appearance reconstruction often receive incorrect object labels during 2D-to-3D projection. We propose OP2GS, an object-aware Gaussian representation that augments each primitive with an explicit instance identity and a dedicated instance opacity

σ^{*}

for object-mask rendering. The original opacity

σ

remains responsible for visual reconstruction, while

σ^{*}

models whether a Gaussian should contribute to a particular object mask. This dual-opacity formulation decouples visual existence from instance occupancy: mislabeled Gaussians can remain available for image rendering while becoming transparent in the object-mask branch. To learn this representation, we introduce a random object loss that optimizes the 1D instance occupancy field using the standard transmittance-based visibility of 3DGS. Semantic descriptors are then attached at the object level through multi-view aggregation, eliminating per-Gaussian feature storage. Compared with feature-training approaches, OP2GS achieves competitive open-vocabulary performance while significantly reducing computational overhead. Compared with training-free pipelines, it leverages physically consistent occupancy learning to resolve visibility ambiguities.

Abstract

3D Gaussian Splatting

Explore similar work

OP2GS: Object-Aware 3D Gaussian Splatting with Dual-Opacity Primitives

Learning Representations from 3D Gaussian Splats

Drop-In Perceptual Optimization for 3D Gaussian Splatting