3D-VITAL: Visibility-aware Identity Training from Artificial Liftings of 3D Reconstruction model for Cross-View Object Re-Identification

William Grolleau; Astrid Sabourin; Guillaume Lapouge; Catherine Achard

cs.CVOct 12, 2026

3D-VITAL: Visibility-aware Identity Training from Artificial Liftings of 3D Reconstruction model for Cross-View Object Re-Identification

Authors: William Grolleau, Astrid Sabourin, Guillaume Lapouge, Catherine Achard

Organizations: Université Paris-Saclay, CEA, List, F-91120 Palaiseau, France · Sorbonne University, CNRS, Institute of Intelligent Systems and Robotics (ISIR), Paris, France

Abstract

We address the Cross-View Object Re-Identification (ReID) task, aiming to match identities across drastically different viewpoints. We focus on Single-View Aerial-Ground ReID (SV AG-ReID), where models train on one viewpoint and must generalize to another at inference. To bridge this viewpoint gap, related work employs generative models to synthesize novel views from source images as additional training data. We leverage recent advances in large-scale single-view 3D mesh reconstruction to lift source images into textured meshes and to render them from new viewpoints. However, this pipeline produces unreliable data in two distinct ways. First, the reconstruction may fail and produce geometrically incorrect meshes, resulting in severe corruption in rendered novel views. Second, even when the geometry is correct, the rendered novel views contain hallucinated textures in regions not visible in the source image. The reconstructed mesh carries a per-pixel partition between regions visible from the source view and regions that the reconstruction model had to hallucinate, available as a deterministic byproduct of rendering. We introduce this per-pixel visibility map and exploit it in a two-stage design. Cross-View Consistency Filtering (CVCF) discards geometrically incorrect meshes by ensuring multi-view consistency between source and novel views over visible patches. Visibility-Aware Attention Supervision (VAAS) leverages the visibility map to guide the model's attention toward reliable regions of the rendered view during training. Combined, these two components enable our method, 3D-VITAL, to better leverage the synthetic signal from single-view 3D mesh reconstruction. 3D-VITAL sets a new state-of-the-art on three SV AG-ReID benchmarks. Code and data will be released at https://github.com/TurtleSmoke/3D-VITAL.

Figures & tables

Figure 1: Motivation of 3D-VITAL. Left: Prior synthetic-real ReID pipelines such as 3D-LENS treat each synthetic sample as wholly reliable or wholly unreliable, which lets misleading identity signal from hallucinated regions (red rectangle) degrade training. Right: 3D-VITAL instead derives a per-pixel visibility map (visible regions as green, hallucinated as red) from the reconstructed mesh, separating observed regions from hallucinated ones. Cross-View Consistency Filtering (CVCF) uses this map to reject geometrically incorrect samples. Visibility-Aware Attention Supervision (VAAS) uses it to guide the model’s attention toward observed regions during training.

Panel of Figure 2 — Figure 2: Cross-View Consistency Filtering (CVCF). (a) Frozen DINOv2 features match each patch of a style-aligned novel view $I_{i}^{\text{sty}}$ to its most similar source-image patch, giving cosine similarities $\sigma_{i}$ ( Equation 1 ). The patch map $\tilde{V}_{i}$ masks these similarities to source-visible regions ( Equation 2 ), so hallucinated patches do not contribute. The visible-patch mean, averaged over the $N$ novel views, gives one consistency score $\bar{s}_{i}$ per source reconstruction ( Equation 3 ), used to keep the top samples. (b) A correct mesh (top) stays consistent across views in the visible regions, even when hallucinated regions are large, while an incorrect mesh (bottom) is inconsistent.

Figure 3: Visibility-Aware Attention Supervision (VAAS). A style-aligned novel view $I_{i}^{\text{sty}}$ is processed by a ViT. At the final transformer layer, we extract the attention weights $\alpha_{i}$ of the identity token $t_{\text{id}}$ over the image patches, averaged across heads. The visibility map $\tilde{V}_{i}$ provides two supervision signals. Top: $\mathcal{L}_{\text{vis}}$ is the cross-entropy between $\alpha_{i}$ and the normalized visibility target $\hat{V}_{i}$ ( Equation 4 ), pulling identity-token attention toward visible patches. Bottom: the per-sample weight $w_{i}=\tfrac{1}{P}\sum_{j}\tilde{V}_{i}^{(j)}$ scales the anchor’s contribution to $\mathcal{L}_{\text{tri}}^{w}$ ( Equation 5 ), reducing the influence of heavily hallucinated samples.

		AG-ReID								MOO
		Aerial Training				Ground Training				Top Training				Side Training
Methods	Year	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$		$T\rightarrow S$		$S\rightarrow T$		$T\rightarrow S$		$S\rightarrow T$
		mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
BoT Luo et al. (2019)	2019	41.7	55.7	29.6	34.3	45.3	51.8	49.6	60.2	37.1	95.9	39.1	43.5	19.3	43.1	19.4	94.2
ViT Dosovitskiy et al. (2021)	2020	42.6	58.1	31.9	35.8	46.9	54.6	49.8	60.1	36.5	93.9	37.8	41.6	19.8	43.5	20.0	94.8
AGW Ye et al. (2022)	2021	41.4	55.3	32.2	36.7	42.3	51.0	47.3	59.5	36.1	89.2	37.8	44.0	13.5	40.8	13.8	76.6

Table 1: State-of-the-Art Comparison on AG-ReID and MOO. mAP and Rank-1 (%) on cross-view scenarios across Aerial (A), Ground (G) for AG-ReID, and Top (T), and Side (S) for MOO. Best in bold , second-best underlined .

AG-ReID.v2
		Aerial Training								CCTV Training				Wearable Training
Methods	Year	$C\rightarrow A$		$A\rightarrow C$		$W\rightarrow A$		$A\rightarrow W$		$C\rightarrow A$		$A\rightarrow C$		$W\rightarrow A$		$A\rightarrow W$
		mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
BoT Luo et al. (2019)	2019	14.4	16.8	32.5	42.4	18.8	21.2	28.7	40.5	42.3	52.6	34.5	40.0	31.6	38.0	25.1	28.9
ViT Dosovitskiy et al. (2021)	2020	14.7	16.7	33.2	43.1	19.1	21.5	27.4	38.1	41.8	51.8	35.0	41.0	31.0	37.1	23.9	27.9
AGW Ye et al. (2022)	2021	14.6	17.8	33.2	42.6	19.9	23.9	30.2	40.9	38.3	47.4	31.5	38.2	32.4	39.1	25.3	31.2

Table 2: State-of-the-Art Comparison on AG-ReID.v2. mAP and Rank-1 (%) on cross-view scenarios across Aerial (A), CCTV (C), and Wearable (W). Best in bold , second-best underlined .

Index	Components			AG-ReID Aerial				AG-ReID Ground
	$\text{VAAS}_{\text{vis}}$	$\text{VAAS}_{\text{tri}}$	CVCF	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
	$\text{VAAS}_{\text{vis}}$	$\text{VAAS}_{\text{tri}}$	CVCF	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
0				45.6	61.2	46.1	56.4	49.7	61.6	51.3	62.9
1	✓			48.8 ( +3.2 )	63.9 ( +2.6 )	50.0 ( +3.8 )	59.6 ( +3.1 )	52.2 ( +2.5 )	63.3 ( +1.7 )	54.6 ( +3.3 )	65.8 ( +2.9 )
2		✓		46.5 ( +0.9 )	62.2 ( +0.9 )	46.9 ( +0.8 )	56.9 ( +0.4 )	50.6 ( +0.9 )	62.9 ( +1.3 )	52.9 ( +1.6 )	63.6 ( +0.7 )
3	✓	✓		49.2 ( +3.6 )	64.2 ( +3.0 )	49.5 ( +3.3 )	58.6 ( +2.2 )	53.0 ( +3.3 )	64.5 ( +2.9 )	55.5 ( +4.3 )	67.6 ( +4.7 )

Table 3: Component Ablation on AG-ReID. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial, with gains over Index 0 in green. Best in bold .

		Aerial Training				Ground Training
Setting	Azimuth	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
		mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
Baseline	$\pm 15^{\circ}$	46.5	61.7	47.2	57.2	50.0	62.3	51.1	63.3
	$\pm 30^{\circ}$	46.9	62.6	47.9	57.9	50.6	62.6	52.9	64.3
	$\pm 60^{\circ}$	45.6	61.2	46.1	56.4	49.7	61.6	51.3	62.9
3D-VITAL	$\pm 15^{\circ}$	48.7	63.0	50.0	58.1	54.9	65.3	56.1	67.0

Table 4: Azimuth Range Ablation on AG-ReID. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial. Best in bold , second-best underlined .

		Aerial Training				Ground Training
Setting	Filter	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
		mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
Without VAAS	No filter	45.6	61.2	46.1	56.4	49.7	61.6	51.3	62.9
	CLIP I-I	46.6	61.0	47.1	57.1	49.0	61.6	51.8	64.1
	CLIP I-T	46.9	61.7	47.7	57.8	51.8	63.5	54.8	65.8
	DINOv2 CLS	48.5	64.2	50.4	59.5	52.0	63.3	54.3	65.8

Table 5: Filtering Strategy Ablation on AG-ReID. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial. Best per column within each sub-block in bold , second-best underlined .

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

	Aerial Training				Ground Training
Variant	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
Visibility map (ours)	52.1	66.1	53.2	62.6	56.0	67.0	57.9	68.6
Hallucinated pixels masked	50.3	63.1	46.5	52.4	54.3	63.0	55.3	66.0

Appendix

Table 6: Hallucinated Pixel Masking on AG-ReID. Hallucinated pixels of the rendered views are masked out of the input. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial. Best in bold .

	Aerial Training				Ground Training
Variant	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
$<10\%$ hallucinated	47.6	62.9	48.4	57.6	51.3	62.3	53.5	64.3
$<30\%$ hallucinated	49.1	63.6	51.2	59.6	52.8	64.6	55.3	66.3
$<50\%$ hallucinated	50.9	64.7	51.7	60.9	54.1	66.4	56.9	67.2
$<70\%$ hallucinated	51.5	65.8	52.5	61.3	55.3	65.7	57.6	67.7

Appendix

Table 7: Hallucinated Fraction Ablation on AG-ReID. Only novel views with less than the given fraction of hallucinated pixels are retained. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial. Best in bold , second-best underlined .

	Aerial Training				Ground Training
Variant	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
No noise (ours)	52.1	66.1	53.2	62.6	56.0	67.0	57.9	68.6
$\pm 5^{\circ}$	52.0	66.1	52.9	62.4	55.8	66.9	57.8	68.4
$\pm 10^{\circ}$	51.2	65.8	52.6	61.9	53.4	63.9	56.8	67.7
$\pm 20^{\circ}$	49.7	63.9	50.6	60.1	52.1	62.0	54.2	66.8

Appendix

Table 8: Calibration Noise Ablation on AG-ReID. Noise of up to

\pm n^{\circ}

in azimuth and elevation is added to the fitted source pose. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial. Best in bold , second-best underlined .

	Aerial Training				Ground Training
Variant	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
One map	50.5	64.0	51.5	59.4	55.0	66.2	54.4	65.3
Foreground map	50.6	65.1	51.6	59.2	54.8	65.3	53.5	64.9
Hallucinated map	48.2	62.5	50.3	57.8	52.8	63.3	55.3	65.3
Visibility map (ours)	52.1	66.1	53.2	62.6	56.0	67.0	57.9	68.6

Appendix

Table 9: Visibility Mask Source Ablation on AG-ReID. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial. Best in bold , second-best underlined .

	Aerial Training				Ground Training
Variant	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
$L=2$	51.2	64.4	52.3	61.4	55.2	65.4	56.4	66.1
$L=4$	50.1	63.8	51.2	59.2	54.0	64.0	55.7	66.0
$L=8$	50.0	63.8	51.2	59.2	53.8	64.8	55.6	66.4
$L=12$	50.2	63.7	51.0	59.5	53.6	64.7	55.7	65.7

Appendix

Table 10: Number of Supervised Attention Layers Ablation on AG-ReID. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial. Best in bold , second-best underlined .

	Aerial Training				Ground Training
Variant	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
$\lambda_{\text{tri}}{=}0.25$	51.1	59.6	49.7	59.4	54.4	65.7	56.6	66.4
$\lambda_{\text{tri}}{=}0.50$	51.5	65.8	52.9	61.3	55.9	66.9	55.3	66.5
$\lambda_{\text{tri}}{=}0.75$	51.9	64.8	52.9	61.3	55.8	66.9	57.8	68.9
$\lambda_{\text{tri}}{=}1.00$ (ours)	52.1	66.1	53.2	62.6	56.0	67.0	57.9	68.6

Appendix

Table 11: Triplet Loss Weight

\lambda_{\text{tri}}

Ablation on AG-ReID. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial. Best in bold , second-best underlined .

	Aerial Training				Ground Training
Variant	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
$\lambda_{\text{vis}}{=}0.25$	51.3	64.8	51.9	59.5	54.2	65.2	56.5	65.9
$\lambda_{\text{vis}}{=}0.50$	51.8	65.7	52.0	60.6	55.9	66.8	57.5	68.5
$\lambda_{\text{vis}}{=}0.75$	51.9	65.5	52.7	61.1	55.6	66.1	57.4	68.5
$\lambda_{\text{vis}}{=}1.00$ (ours)	52.1	66.1	53.2	62.6	56.0	67.0	57.9	68.6

Appendix

Table 12: CLS-Attention Loss Weight

\lambda_{\text{vis}}

Ablation on AG-ReID. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial. Best in bold , second-best underlined .

	Aerial Training				Ground Training
Variant	$A\rightarrow G$		$G\rightarrow A$		$A\rightarrow G$		$G\rightarrow A$
	mAP	R-1	mAP	R-1	mAP	R-1	mAP	R-1
$0.10$	49.2	62.1	45.9	53.3	53.1	62.5	56.2	68.1
$0.20$	50.6	62.3	51.5	59.5	55.3	66.0	57.1	68.7
$0.30$	51.7	63.9	52.9	61.8	55.3	65.3	57.3	67.2
$0.40$	51.8	66.0	53.0	62.4	55.3	65.9	57.6	67.8

Appendix

Table 13: Filtering Ratio Ablation on AG-ReID. Percentage of filtered novel meshes. mAP and Rank-1 (%) on Aerial

\rightarrow

Ground and Ground

\rightarrow

Aerial. Best in bold , second-best underlined .

Person Re-Identification

Person re-identification matches the same individual across camera views, times, clothing, and sensing modalities despite changes in appearance.

See all 90 papers

Apr 29, 2026

3D-LENS: A 3D Lifting-based Elevated Novel-view Synthesis method for Single-View Aerial-Ground Re-Identification

Aerial-Ground Re-Identification (AG-ReID) is constrained by the viewpoint-domain gap, as drastic viewpoint disparities occlude or distort discriminative features, making cross-viewpoint image retrieval challenging. While existing methods rely on paired cross-view annotations, real-world deployments, such as wilderness search-and-rescue (SAR), often lack target-domain data, requiring retrieval from ground-level references alone. To our knowledge, we are the first to address this challenge by formalizing the Single-View AG-ReID (SV AG-ReID) setting, where models trained on a single real viewpoint must generalize to an unseen viewpoint. We propose 3D Lifting-based Elevated Novel-view Synthesis (3D-LENS), a unified framework combining geometrically-consistent novel view synthesis that leverages large-scale 3D mesh reconstruction, with a robust representation learning scheme to mitigate synthetic-to-real bias. Unlike 2D generative baselines that suffer from geometric inconsistencies or prior 3D methods that are restricted to class-specific templates, our approach ensures view-consistent synthesis across diverse categories without predefined templates that fail to capture fine-grained details, such as carried objects. Extensive experiments demonstrate that our method achieves state-of-the-art performance on SV AG-ReID scenarios. Code and data will be released at https://github.com/TurtleSmoke/3D-LENS.

William Grolleau, Astrid Sabourin, Guillaume Lapouge +1

Université Paris-Saclay, CEA, List, F-91120 Palaiseau, France · Sorbonne University, CNRS, Institute of Intelligent Systems and Robotics (ISIR), 4 Place Jussieu, 75005 Paris, France

Novel View Synthesis Person Re-Identification

Ask PDF

Aug 3, 2026

VR3D: View-Robust 3D Representation Learning for Aerial-Ground Person Re-Identification

Aerial-ground person re-identification is a challenging task due to cross-platform viewpoint variations, which cause severe occlusion and geometric deformation. Existing methods attempt to learn view-invariant representations exclusively within the 2D image space, where drastic viewpoint variations cause the learned features to remain coupled with viewpoint bias. To address this, we propose VR3D, a View-Robust 3D Representation Learning framework that maps images into a unified 3D coordinate space to achieve view-independent feature interaction. Specifically, we introduce View-Robust 3D Representation Interaction, which leverages 3D priors extracted from single 2D observations to lift 2D appearance features into a canonical 3D space. VR3I employs 3D Geometry-Semantic Attention to establish interactions between 2D patches and 3D voxels from corresponding body parts based on their 3D spatial locations, effectively grounding 2D semantics within a 3D framework. In addition, as the reliability of these representations varies across samples due to viewpoint changes and 3D reconstruction errors, we introduce Reliability-Aware Fusion, which estimates sample-specific reliability and adaptively aggregates the multi-source representations. Extensive experiments on three benchmark datasets (CARGO, AG-ReID.v1, and AG-ReID.v2) demonstrate that VR3D outperforms recent methods. For example, it achieves a 5.63% improvement in Rank-1 on CARGO. Our code will be released.

Chao Ji, Shiyu Xuan, Zechao Li

School of Computer Science and Engineering, Nanjing University of Science and Technology

3D Representation Learning Invariant Representation Learning Person Re-Identification

Ask PDF

May 18, 2026

View-Aware Semantic Alignment for Aerial-Ground Person Re-Identification

Aerial-Ground Person Re-Identification (AGPReID) remains highly challenging due to drastic viewpoint variations between drones and fixed cameras. Existing methods typically follow a view-invariant paradigm, aligning shared features across views to achieve robustness. However, view-invariant inherently enforces part-level alignment, which ignores view-specific cues and discriminative identity information. To this end, this work proposes ViSA (View-aware Semantic Alignment), a view-aware framework that achieves cross-view semantic consistency containing an Expert-driven Token Generation Module (ETGM) and a Dual-branch Local Fusion Module (DLFM). Technically, the former constructs a set of view-aware experts to generate adaptive semantic queries that perceive viewpoint-specific patterns, while the latter leverages graph reasoning to extract and align local regions responsive to different experts. Extensive experiments on three AGPReID benchmarks including AG-ReID.v2, CARGO and LAGPeR demonstrate that ViSA consistently achieves superior performance, with a notable 10.06% mAP improvement on the challenging CARGO cross-view protocol. The code is available at https://github.com/Cat-Zero/ViSA.

Quan Zhang, Zeqiang Cai, Peiming Zhao +4

Sun Yat-sen University, China · Pazhou Lab (HuangPu), Guangdong, China · Guangdong Province Key Laboratory of Information Security Technology, Guangzhou, China +1

Multi-View Learning Person Re-Identification

Ask PDF

Abstract

Person Re-Identification

Explore similar work

3D-LENS: A 3D Lifting-based Elevated Novel-view Synthesis method for Single-View Aerial-Ground Re-Identification

VR3D: View-Robust 3D Representation Learning for Aerial-Ground Person Re-Identification

View-Aware Semantic Alignment for Aerial-Ground Person Re-Identification