TCC-VPReID: Three-Stream Person Re-ID

• The paper introduces a three-stream architecture that integrates style-robust feature learning, memory-based cross-view adaptation, and intermediary-guided temporal learning for person re-ID.
• It fuses sequence-level descriptors from different modalities and platforms to effectively address variations in viewpoint, scale, and illumination.
• Empirical results demonstrate notable performance improvements with enhanced Rank-1 and mAP scores on challenging aerial–ground video matching protocols.

Searching arXiv for the primary paper and a few directly mentioned related methods to ground the article in current sources. TCC‑VPReID, short for Three‑stream architecture for Cross‑platform Cross‑modality Video‑based Person Re‑ID, is a person re‑identification framework proposed for the joint setting of video-based, cross‑modality, and cross‑platform matching: tracklets rather than still images, visible (RGB) versus infrared (IR/thermal) imagery, and aerial (UAV) versus ground (CCTV / wearable) acquisition (Nguyen et al., 24 Jul 2025). It is introduced together with AG‑VPReID.VIR, described as the first aerial‑ground cross‑modality video‑based person Re‑ID dataset, and is designed to address the entangled gaps induced by viewpoint, platform, modality, scale, and time. In the reported formulation, TCC‑VPReID combines style‑robust feature learning, memory‑based cross‑view adaptation, and intermediary‑guided temporal learning, then fuses the resulting sequence‑level descriptors for distance‑based matching (Nguyen et al., 24 Jul 2025).

1. Problem setting and dataset context

TCC‑VPReID is defined for what the source paper describes as its most challenging Re‑ID configuration: video-based inputs as tracklets, cross‑modality matching between RGB and IR, and cross‑platform matching between aerial and ground cameras (Nguyen et al., 24 Jul 2025). This formulation is tied directly to the construction of AG‑VPReID.VIR, which contains 1,837 identities, 4,861 tracklets, and 124,855 frames collected with UAV RGB and UAV IR (altitudes 15–45m), CCTV RGB and CCTV IR (≈3–4m height), and wearable RGB at ≈1.5m (Nguyen et al., 24 Jul 2025).

The dataset design creates several combined difficulties. The paper enumerates a viewpoint / platform gap between aerial top‑down and ground side views, with large camera height differences and extreme pose changes; scale & resolution variation, with people becoming very small in aerial views, down to 31×59 px; a modality gap, since RGB encodes color and texture whereas IR encodes heat patterns and is grayscale‑like and often low contrast, especially for aerial IR; and temporal dynamics such as motion blur, temporal appearance changes, partial occlusions, group interactions, and background clutter (Nguyen et al., 24 Jul 2025).

Within that setting, TCC‑VPReID is not presented as a generic video Re‑ID backbone. It is instead tailored to a joint aerial‑ground and visible‑infrared regime. This suggests that its three‑stream factorization is meant to separate partially distinct nuisance sources—style, viewpoint, and modality‑temporal structure—while still permitting end‑to‑end optimization on fused sequence representations.

2. Three‑stream formulation and processing pipeline

The architecture is explicitly described as a three‑stream architecture composed of: Stream 1: Style‑Robust Feature Learning; Stream 2: Memory‑based Cross‑View Adaptation; and Stream 3: Intermediary‑Guided Temporal Learning (Nguyen et al., 24 Jul 2025). A feature fusion module integrates the three streams into a fused representation used for matching; during training, all streams are optimized jointly, and during inference the fused feature serves as the final descriptor (Nguyen et al., 24 Jul 2025).

Before detailing the individual streams, the overall decomposition can be summarized as follows.

Stream	Primary target	Backbone / core mechanism
Stream 1	Intra / inter‑modal style variations and appearance distortions	ResNet‑50 with style augmentation and style disturbance defense
Stream 2	Aerial–ground viewpoint differences and temporal cross‑view inconsistency	CLIP ViT‑B/16 visual encoder + 2‑layer Transformer decoder + view‑specific identity memories
Stream 3	RGB–IR modality gap and temporal aggregation	Dual‑branch network with 3D cross‑attention + Bi‑LSTM, using anaglyph intermediaries

For each tracklet from any camera and modality, the pipeline proceeds in four stages (Nguyen et al., 24 Jul 2025). First, at frame level, RGB or IR frames are fed into each stream’s backbone; Stream 1 processes style‑augmented frames, while Stream 3 converts frames into anaglyph (edge) images. Second, at temporal level, each stream aggregates frame‑wise features into a sequence‑level feature $v_a$, and Stream 2 additionally uses a Transformer to encode temporal context and build or update memory. Third, at the adaptation stage, Stream 2 aligns sequence features to aerial / ground identity memories, while Stream 3 enforces cross‑modality reconstruction between RGB and IR anaglyph features. Fourth, at fusion and matching, the three sequence‑level features are fused and the fused tracklet descriptor is used for distance‑based matching between query and gallery (Nguyen et al., 24 Jul 2025).

The paper states that fusion uses a “feature fusion module” with adaptive weighting, but does not fully specify the fusion strategy. It also states that the exact implementation may involve adaptive weighting based on validation performance (Nguyen et al., 24 Jul 2025). Accordingly, the fused representation is conceptually central, while the low‑level fusion operator remains only partially specified in the published description.

3. Style‑robust feature learning stream

Stream 1 is designed to handle intra / inter‑modal style variations and appearance distortions across aerial versus ground and RGB versus IR, with the explicit goal of learning identity features robust to illumination, color, spectral differences, viewpoint, scale, and other appearance distortions (Nguyen et al., 24 Jul 2025). Its backbone is ResNet‑50, described as shared across modalities and platforms, with style augmentation applied to the input stream (Nguyen et al., 24 Jul 2025).

The first mechanism is style augmentation, adopted and extended from video style augmentation attributed to Zhou et al. For an RGB frame $I_t$ with channels $I_t^R, I_t^G, I_t^B$,

$$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$

and for an IR frame $I_t^{ir}$,

$I_t^{ir,aug} = \delta I_t^{ir},$

where

$\alpha, \beta, \gamma, \delta \sim U(0.5, 1.5).$

This introduces synthetic style variations through channel‑wise or scalar intensity changes, encouraging invariance to brightness and contrast perturbations (Nguyen et al., 24 Jul 2025).

The second mechanism is an intra‑modal style attack within the network, inserted after the third convolutional block. Let $f_{conv3}^{i,j}$ denote the conv3 feature map of identity $i$, frame $j$, and let $I_t$0 be a randomly selected feature from another identity‑frame pair. The style‑attacked feature is constructed as

$I_t$1

where $I_t$2 and $I_t$3 are mean and variance over channels (Nguyen et al., 24 Jul 2025). The paper interprets this as replacing the statistics of one sample with those of another while retaining semantic content, in practice akin to Adaptive Instance Normalization.

The associated objective is a style‑attack loss

$I_t$4

The paper states that $I_t$5 enforces identity consistency through cross‑entropy loss and that $I_t$6 is a consistency loss between original and attacked features through the distance between original and attacked features (Nguyen et al., 24 Jul 2025). It provides the discriminative term in cross‑entropy form,

$I_t$7

and describes the consistency term conceptually as a distance between feature heads on original and attacked features (Nguyen et al., 24 Jul 2025).

This stream is also trained jointly with the model‑level identity classification loss $I_t$8 and triplet loss $I_t$9, so that the ResNet features are both identity‑discriminative and style‑robust over aerial versus ground and RGB versus IR (Nguyen et al., 24 Jul 2025). The supplementary material is said to hint at graph interactions and domain‑adversarial alignment, but these are only high‑level descriptions and are not detailed mathematically in the paper (Nguyen et al., 24 Jul 2025).

4. Memory‑based cross‑view adaptation stream

Stream 2 addresses aerial versus ground viewpoint variations and temporal cross‑view inconsistency by building view‑specific identity memories and aligning sequence features with them via contrastive learning (Nguyen et al., 24 Jul 2025). Its backbone is a CLIP ViT‑B/16 visual encoder followed by a 2‑layer Transformer decoder that models temporal relationships and produces a sequence‑level feature vector $I_t^R, I_t^G, I_t^B$0 (Nguyen et al., 24 Jul 2025).

For each identity $I_t^R, I_t^G, I_t^B$1, the stream constructs two prototype memories: $I_t^R, I_t^G, I_t^B$2 These are defined as averages over all aerial or ground sequences of that identity: $I_t^R, I_t^G, I_t^B$3

$I_t^R, I_t^G, I_t^B$4

where $I_t^R, I_t^G, I_t^B$5 and $I_t^R, I_t^G, I_t^B$6 (Nguyen et al., 24 Jul 2025). In matrix form, the two memory banks are

$I_t^R, I_t^G, I_t^B$7

The paper further introduces a cross‑platform attention mechanism that updates each memory using a platform‑specific prompt: $I_t^R, I_t^G, I_t^B$8 Here $I_t^R, I_t^G, I_t^B$9 is produced by a dual‑branch attention‑based decoder and is intended to encourage cross‑platform information sharing while preserving platform‑unique characteristics (Nguyen et al., 24 Jul 2025). The detailed attention equations for $$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$0 are not explicitly given.

Alignment is supervised through a video‑to‑memory contrastive loss $$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$1: $$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$2 where $$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$3 is the set of positive samples of identity $$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$4 in the batch, $$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$5 is the batch size, and $$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$6 is a temperature parameter (Nguyen et al., 24 Jul 2025). The effect is InfoNCE‑like: each sequence feature is pulled toward its identity‑and‑platform specific memory and pushed away from other batch sequences.

The architectural rationale is explicit. By maintaining separate aerial and ground memories rather than a single shared prototype, Stream 2 learns two coupled manifolds for each identity and aligns them via prompts (Nguyen et al., 24 Jul 2025). The paper reports that this stream is the strongest individual component in ablation, which is consistent with the claim that cross‑platform viewpoint variation is a core source of difficulty on AG‑VPReID.VIR.

5. Intermediary‑guided temporal learning stream

Stream 3 is devoted to the RGB–IR modality gap and temporal aggregation, using modality‑invariant intermediary representations derived from edges (Nguyen et al., 24 Jul 2025). Following IBAN, it constructs anaglyph images from both RGB and IR inputs through an edge operator: $$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$7 where $$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$8 is the $$I_t^{aug} = [\alpha I_t^R, \beta I_t^G, \gamma I_t^B],$$9-th pixel of the anaglyph, $I_t^{ir}$0 is an edge detection operator with $I_t^{ir}$1, and $I_t^{ir}$2 is an offset (Nguyen et al., 24 Jul 2025).

The paper’s stated intuition is that edges and gradients are more modality‑invariant than raw pixels: silhouettes, contours, and motion boundaries appear more similar across RGB and IR than color or thermal texture (Nguyen et al., 24 Jul 2025). The anaglyph images are therefore used as intermediaries that suppress modality‑specific noise and foreground structural cues useful for identity matching.

These representations are processed by a dual‑branch network in which one branch handles visible anaglyphs $I_t^{ir}$3 and the other IR anaglyphs $I_t^{ir}$4. The network includes spatial‑temporal 3D cross‑attention and bidirectional LSTMs for sequence modeling (Nguyen et al., 24 Jul 2025). While the paper does not detail the 3D cross‑attention equations, it does describe standard Bi‑LSTM temporal aggregation over frame‑level features.

To enforce modality invariance, Stream 3 introduces a cross‑reconstruction network $I_t^{ir}$5 and a cross‑reconstruction loss

$I_t^{ir}$6

where $I_t^{ir}$7 is the mini‑batch size and $I_t^{ir}$8 the number of frames per sequence (Nguyen et al., 24 Jul 2025). This bidirectionally constrains $I_t^{ir}$9 and $I_t^{ir,aug} = \delta I_t^{ir},$0, making the two modality representations more predictable from each other.

The paper uses the term “intermediary‑guided” to indicate that the anaglyph representations and their reconstructed versions serve as anchors for temporal aggregation, guiding the LSTM toward modality‑invariant patterns such as edges and motion contours (Nguyen et al., 24 Jul 2025). Stream 3 alone is weaker than Stream 2 in cross‑platform settings, but the reported ablations indicate that it contributes complementary gains when combined with the other streams.

6. Joint optimization, inference, and empirical behavior

The overall training objective combines identity supervision, metric learning, style robustness, modality alignment, and memory adaptation: $I_t^{ir,aug} = \delta I_t^{ir},$1 The paper identifies the best hyperparameter configuration in its ablation study as

$I_t^{ir,aug} = \delta I_t^{ir},$2

(Nguyen et al., 24 Jul 2025). It interprets this as evidence that style robustness and memory adaptation are particularly important on the hardest protocol, Aerial→Ground I2V (Nguyen et al., 24 Jul 2025).

At test time, each query and gallery tracklet is processed by all three streams, a sequence‑level feature is obtained from each, and the features are fused into a single tracklet descriptor (Nguyen et al., 24 Jul 2025). The exact fusion implementation is not explicitly specified; the paper states only that adaptive weighting is applied to each stream’s contribution based on validation performance (Nguyen et al., 24 Jul 2025). The descriptor is then used for distance‑based ranking, and the paper reports results using CMC and mAP metrics (Nguyen et al., 24 Jul 2025).

The reported results establish several empirical patterns. On ground↔ground cross‑modality evaluation on AG‑VPReID.VIR, TCC‑VPReID achieves Rank‑1 = 36.18%, mAP = 41.56 for I2V and Rank‑1 = 46.33%, mAP = 59.23 for V2I (Nguyen et al., 24 Jul 2025). On cross‑platform protocols, the full three‑stream model (St123) reports Aerial→Ground I2V at Rank‑1 = 19.83%, Rank‑5 = 31.90, Rank‑10 = 43.10, Rank‑20 = 51.72, mAP = 22.61, and Aerial→Ground V2I at Rank‑1 = 46.54%, Rank‑5 = 76.54, Rank‑10 = 85.38, Rank‑20 = 91.92, mAP = 59.69 (Nguyen et al., 24 Jul 2025). On aerial↔aerial cross‑modality evaluation, St123 attains RGB→IR Rank‑1 = 33.91, mAP = 44.72 and IR→RGB Rank‑1 = 21.84, mAP = 30.92 (Nguyen et al., 24 Jul 2025).

The ablation evidence is equally central to understanding the method. Stream 2 alone is reported as the strongest individual stream, with Aerial→Ground V2I: Rank‑1 = 40.38%, mAP = 54.40 and Aerial→Ground I2V: Rank‑1 = 11.21%, mAP = 17.68 (Nguyen et al., 24 Jul 2025). Stream 1 alone is second, and the combination St12 improves over Stream 2 alone in Aerial→Ground I2V from 11.21 to 15.52% Rank‑1, while the full St123 yields the best overall results (Nguyen et al., 24 Jul 2025). The paper interprets this to mean that cross‑view memory is core to handling aerial–ground variation, while style robustness and modality‑invariant temporal modeling are complementary (Nguyen et al., 24 Jul 2025).

The same paper also states that TCC‑VPReID surpasses prior state of the art on HITSZ‑VCM and BUPTCampus, both described as ground‑only RGB–IR video datasets (Nguyen et al., 24 Jul 2025). That claim situates the framework beyond the aerial‑ground benchmark on which it is introduced, although the article’s defining use case remains AG‑VPReID.VIR.

7. Scope, unresolved details, and limitations

Several aspects of TCC‑VPReID are explicit strengths of the formulation, while others remain partially specified in the published presentation. The architecture clearly separates style invariance, cross‑platform prototype alignment, and modality‑invariant temporal modeling, and trains them with a joint loss over fused sequence embeddings (Nguyen et al., 24 Jul 2025). At the same time, the paper states that the fusion strategy is not fully specified, the distance metric at inference is not explicitly stated, and some mechanisms alluded to in supplementary descriptions—such as graph interactions, domain‑adversarial alignment, and temporal memory diffusion—are only described at a high level (Nguyen et al., 24 Jul 2025). For encyclopedia purposes, these are best treated as components whose conceptual roles are clear but whose exact implementation details are incomplete in the public mathematical exposition.

The limitations discussed or inferable from the paper’s own results are also specific. Although TCC‑VPReID improves robustness, performance on Aerial→Ground I2V remains relatively modest, which the source links to very low resolution / extreme altitudes and hard aerial IR conditions (Nguyen et al., 24 Jul 2025). The dataset includes occluded and group scenes, and the paper notes remaining failure cases in crowded scenarios (Nguyen et al., 24 Jul 2025). It also does not evaluate cross‑dataset or domain generalization, since training and testing occur within the same campus domain (Nguyen et al., 24 Jul 2025).

Within the literature of person re‑identification, TCC‑VPReID is therefore best understood as a specialized framework for a compound regime that earlier RGB–IR video Re‑ID benchmarks did not directly target: simultaneous aerial–ground, visible–infrared, and video‑based matching (Nguyen et al., 24 Jul 2025). Its defining contribution is the explicit decomposition of that regime into three coupled learning problems—style invariance, view‑specific memory alignment, and intermediary‑guided temporal modeling—and the empirical claim that their fusion yields the strongest performance across ground↔ground, aerial↔aerial, and aerial↔ground protocols on AG‑VPReID.VIR (Nguyen et al., 24 Jul 2025).