Scale-Aware Vision-Language Adaptation for Extreme Far-Distance Video Person Re-identification

Ashwat Rajbhandari; Bharatesh Chakravarthi

cs.CVApr 5, 2026

Scale-Aware Vision-Language Adaptation for Extreme Far-Distance Video Person Re-identification

Authors: Ashwat Rajbhandari, Bharatesh Chakravarthi

Organizations: Arizona State University

Abstract

Extreme far-distance video person re-identification (ReID) is particularly challenging due to scale compression, resolution degradation, motion blur, and aerial-ground viewpoint mismatch. As camera altitude and subject distance increase, models trained on close-range imagery degrade significantly. In this work, we investigate how large-scale vision-language models can be adapted to operate reliably under these conditions. Starting from a CLIP-based baseline, we upgrade the visual backbone from ViT-B/16 to ViT-L/14 and introduce backbone-aware selective fine-tuning to stabilize adaptation of the larger transformer. To address noisy and low-resolution tracklets, we incorporate a lightweight temporal attention pooling mechanism that suppresses degraded frames and emphasizes informative observations. We retain adapter-based and prompt-conditioned cross-view learning to mitigate aerial-ground domain shifts, and further refine retrieval using improved optimization and k-reciprocal re-ranking. Experiments on the DetReIDX stress-test benchmark show that our approach achieves mAP scores of 46.69 (A2G), 41.23 (G2A), and 22.98 (A2A), corresponding to an overall mAP of 35.73. These results show that large-scale vision-language backbones, when combined with stability-focused adaptation, significantly enhance robustness in extreme far-distance video person ReID.

Figures & tables

Figure 1: Illustration of extreme far-distance conditions in the DetReIDX dataset. As UAV altitude and horizontal distance increase, pedestrians undergo severe scale compression and resolution degradation. The red bounding boxes highlight the same group of individuals as their pixel footprint shrinks dramatically, showing the visual challenges of aerial-ground video person ReID.

Dataset	Camera	Format	PIDs	BBoxes	Height (m)	Distance (m)
AG-ReID.v2	UAV+CCTV	Still	1615	100.6K	15-45	–
G2APS-ReID	UAV+CCTV	Still	2788	200.8K	20-60	–
DetReIDX	DSLR+UAV	Video+Still	334	13M	5-120	10-120

Table 1: Comparison of aerial-ground benchmark datasets for person detection, ReID, tracking, and action recognition.

Parameter	Stage 1	Stage 2
Optimizer	Adam	Adam
Base Learning Rate	$3.5\times 10^{-4}$	$1\times 10^{-4}$
Max Epochs	120	120
Weight Decay	$1\times 10^{-4}$	$2.5\times 10^{-4}$
Weight Decay (Bias)	$1\times 10^{-4}$	$1\times 10^{-4}$
Images Per Batch	16	16

Table 2: Baseline training configuration.

Parameter	Setting
Batch Size	16
Sequence Length	16
Feature Normalization	$\ell_{2}$ normalization
Distance Metric	Cosine similarity
Re-ranking	Disabled

Table 3: Baseline inference configuration.

Figure 2: Overview of the proposed CLIP ViT-L/ $14$ -based video ReID framework. Augmented tracklets are encoded by a CLIP backbone enhanced with PBP prompt conditioning and adapter modules, with only the final two transformer blocks unfrozen during fine-tuning. Frame-level features are aggregated via temporal attention pooling and optimized using identity and triplet losses.

Stage 1 Training
Parameter	Baseline (ViT-B/16)	Ours (ViT-L/14)
Backbone	ViT-B/16	ViT-L/14
Stride Size	16	14
PBP Prompt Length	1	4
QATW Module	Disabled	Enabled
Instance Norm	Disabled	Enabled

Table 4: Optimization refinements compared to the baseline configuration.

Method	A $\rightarrow$ A				A $\rightarrow$ G				G $\rightarrow$ A
Method	R1	R5	R10	mAP	R1	R5	R10	mAP	R1	R5	R10	mAP
VSLA [ 21 ]	15.96	26.10	32.77	13.83	28.96	54.71	69.13	41.63	58.43	65.17	69.66	26.26
SINet [ 22 ]	14.06	24.51	30.94	12.85	25.62	52.47	66.57	38.46	23.50	49.44	59.55	16.98
PSTA [ 23 ]	13.00	23.80	30.30	10.50	22.30	46.90	59.70	34.40	40.40	56.20	59.60	17.00
BiCNet-TKS [ 24 ]	13.30	26.78	36.57	9.71	21.71	44.85	59.30	33.28	41.57	58.43	65.17	22.12
DUT_IIAU_LAB [ 25 ]	25.39	39.58	48.44	20.13	37.77	65.26	75.31	43.93	69.66	80.90	87.64	35.44

Table 5: Comparison of baseline methods and our method on VReID-XFD.

Figure 3: mAP comparison across baseline methods, the best public result, and our method on the DetReIDX benchmark.

Variant	Backbone	mAP (%)	$\Delta$
CLIP ViT-B/16	ViT-B/16	28.11	+0.00
+ Optimization		28.18	+0.07
+ k-reciprocal re-ranking		29.40	+1.29
+ Re-ranking tuned		29.47	+1.36
+ Temporal attention pooling		29.71	+1.60
+ Re-ranking tuned (final)		29.84	+1.73

Table 6: Ablation study on DetReIDX using Eq. 5.

Person Re-Identification

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

See all 92 papers

ALADIN:Attribute-Language Distillation Network for Person Re-Identification
Mar 23, 2026Wang Zhou, Boran Duan, Haojun Ai +2Person Re-Identification Cross-Modal Knowledge Distillation
Wuhan University No.299 Bayi Road, 430072, PR China
Prompt-Anchored Vision-Text Distillation for Lifelong Person Re-identification
May 6, 2026Wen Wen, Hao Chen, Shiliang ZhangCross-Modal Knowledge Distillation Continual Learning
University of Electronic Science and Technology of China, China · Harbin Institute of Technology, Shenzhen, China · Peking University, China
Multi-Grained Vision-Language Alignment for Domain Generalized Person Re-Identification
Mar 14, 2026Jiachen Li, Xiaojin Gong, Dongping ZhangPerson Re-Identification Vision-Language Alignment
College of Information Science and Electronic Engineering, Zhejiang University · College of Information Engineering, China Jiliang University

Mar 23, 2026J/K move · Enter open · S save

ALADIN:Attribute-Language Distillation Network for Person Re-Identification

Wang Zhou, Boran Duan, Haojun Ai, Ruiqi Lan+1

Wuhan University No.299 Bayi Road, 430072, PR China

Person Re-Identification Cross-Modal Knowledge Distillation Vision-Language Alignment Visual Prompt Learning

Open paper AskPDF arXiv

Recent vision-language models such as CLIP provide strong cross-modal alignment, but current CLIP-guided ReID pipelines rely on global features and fixed prompts. This limits their ability to capture fine-grained attribute cues and adapt to diverse appearances. We propose ALADIN, an attribute-language distillation network that distills knowledge from a frozen CLIP teacher to a lightweight ReID student. ALADIN introduces fine-grained attribute-local alignment to establish adaptive text-visual correspondence and robust representation learning. A Scene-Aware Prompt Generator produces image-specific soft prompts to facilitate adaptive alignment. Attribute-local distillation enforces consistency between textual attributes and local visual features, significantly enhancing robustness under occlusions. Furthermore, we employ cross-modal contrastive and relation distillation to preserve the inherent structural relationships among attributes. To provide precise supervision, we leverage Multimodal LLMs to generate structured attribute descriptions, which are then converted into localized attention maps via CLIP. At inference, only the student is used. Experiments on Market-1501, DukeMTMC-reID, and MSMT17 show improvements over CNN-, Transformer-, and CLIP-based methods, with better generalization and interpretability.

Abstract

Person Re-Identification

Explore similar work

ALADIN:Attribute-Language Distillation Network for Person Re-Identification

Prompt-Anchored Vision-Text Distillation for Lifelong Person Re-identification

Multi-Grained Vision-Language Alignment for Domain Generalized Person Re-Identification