Awiros/person-attribute-recognition

Awiros Person Attribute Recognition: Data-Efficient ConvNeXt V2 Training and a Real-World Benchmark Dataset

Awiros

Abstract

Person attribute recognition is a useful primitive for video analytics, search, retrieval, and event understanding, but production performance is difficult to obtain from conventional supervised training alone. Person crops observed in deployed camera systems vary substantially in scale, illumination, viewpoint, occlusion, image quality, and indoor-versus-outdoor capture conditions, while obtaining exhaustive and reliable attribute annotations at comparable scale is expensive.

This work presents two releases: the Awiros Person Attribute Recognition model, a compact ConvNeXt V2 Tiny multi-head classifier, and the Awiros Person Attribute Benchmark Dataset, a gated evaluation resource containing 35,831 real-world person crops covering 17 attributes related to appearance, clothing, carried objects, accessories, visibility, orientation, occlusion, and image quality. The benchmark contains substantial coverage of both outdoor imagery and indoor/night conditions and is intended to measure attribute recognition under operating conditions representative of deployed camera systems rather than curated web imagery alone.

The model is trained using a two-stage data-efficient pipeline. First, the ConvNeXt backbone is pretrained on a substantially larger corpus of unlabeled person crops using a SimCLR-style contrastive objective together with feature alignment to frozen DINOv2 representations. This stage allows the compact backbone to learn invariances from deployment-domain imagery while being regularized toward the semantic structure captured by a larger self-supervised representation model. Second, supervised examples are generated using a general-purpose vision-language model. Two independently constructed prompts are evaluated for each candidate label, and a pseudo-label is retained only when both prompts produce the same prediction. These consensus-filtered labels are then used to train the attribute heads.

Development experiments showed that the representation-pretraining stage was important for reaching the reported operating point; supervised-only training did not provide the same robustness. On the released benchmark, the current Awiros ConvNeXt V2 Tiny model achieves 82.18% mean and 82.75% median top-1 accuracy across the 17 attribute heads. Under the same benchmark protocol, this exceeds the aggregate performance of the evaluated frozen-representation probes based on PE-Core-G14 and DINOv3 ViT-7B/16.

Awiros person-attribute models from this lineage have been deployed in production since mid-2024. The current model demonstrates that a compact specialist trained with large-scale unlabeled data, representation alignment, and carefully filtered pseudo-supervision can remain competitive with substantially larger general-purpose visual representations while retaining a production-oriented inference architecture.

1. Introduction

Person attribute recognition (PAR) aims to infer structured visual properties from a cropped image of a person. Depending on the application, these properties may describe clothing, accessories, carried objects, body visibility, orientation, occlusion, image quality, or other observable characteristics. Unlike identity recognition, the task does not require determining who a person is; instead, it converts visual appearance into a set of independently useful semantic attributes.

In real camera deployments, PAR is considerably more difficult than evaluation on clean, curated imagery suggests. Person crops may occupy only a small portion of the source frame. Subjects may be partially occluded, viewed from behind or from the side, motion blurred, poorly illuminated, or captured by cameras with different resolutions and compression characteristics. Night scenes introduce an additional appearance shift, while indoor cameras differ from outdoor cameras in lighting, scale, viewing geometry, and background statistics.

A second difficulty is data. Modern visual representation models can learn useful structure from extremely large unlabeled corpora, but domain-specific person-attribute labels are expensive to produce at the same scale. Manual annotation across many attributes further compounds the problem: one crop may require multiple independent judgments, some attributes may be visually ambiguous, and annotation quality can vary substantially between labelers.

A straightforward supervised approach therefore leaves much of the available deployment data unused. Conversely, directly deploying a large general-purpose vision model is not always attractive for an always-on video analytics pipeline, where inference cost, memory footprint, deterministic execution, deployment format, and throughput matter alongside accuracy.

We approach the problem as one of representation learning followed by label-efficient specialization. The central idea is to use large volumes of unlabeled domain imagery for the expensive part of representation formation, and to reserve noisy machine-generated labels for the comparatively smaller task of learning attribute decision boundaries.

The principal contributions of this work are:

1. A real-world person-attribute benchmark. We release a gated benchmark of 35,831 person crops spanning 17 attribute heads and diverse outdoor, indoor, night, occlusion, viewpoint, and image-quality conditions.

2. Unlabeled domain pretraining for a compact backbone. ConvNeXt V2 Tiny is trained on a substantially larger collection of unlabeled person crops using a SimCLR-style contrastive objective.

3. Semantic alignment with DINOv2. In addition to instance-level contrastive learning, the compact representation is explicitly aligned with features from a frozen DINOv2 model, transferring useful visual structure without introducing the teacher into deployment-time inference.

4. Consensus-filtered VLM supervision. Attribute pseudo-labels are generated using two independently constructed prompts. Only prompt pairs that agree are retained, trading label quantity for higher precision.

5. Production and benchmark validation. The resulting compact multi-head model belongs to a production lineage deployed since mid-2024 and obtains the strongest aggregate result among the evaluated systems under the released benchmark protocol.

The broader goal is not to argue that compact specialist models universally outperform vision foundation models. Rather, this work asks a more deployment-oriented question: how much of the representational value of large models can be transferred into a small model when abundant unlabeled domain data is available but reliable human labels are not?

2. Problem Formulation

2.1 Multi-Head Person Attribute Classification

Given a person crop (x), the model predicts a set of attributes

$$\widehat{Y}=\{{\widehat{y}}_1,{\widehat{y}}_2,\dots ,{\widehat{y}}_H\},\backslash hat\{Y\}\; =\; \backslash \{\backslash hat\{y\}\_1,\backslash hat\{y\}\_2,\backslash ldots,\backslash hat\{y\}\_\{H\}\backslash \},$$

where (H=17) is the number of attribute heads.

A shared visual encoder (f_\theta) maps the input image to a common representation,

$$z=f_{\theta }(x),z\; =\; f\_\backslash theta(x),$$

and each attribute is predicted by a dedicated classification head (g_h):

$${\widehat{y}}_h=\arg \underset{c}{\max }{g}_h(z{)}_c\mathrm{.}\backslash hat\{y\}\_h\; =\; \backslash arg\backslash max\_c\; g\_h(z)\_c.$$

This shared-backbone formulation is well suited to person attributes because many tasks depend on overlapping visual cues. Clothing type, carried objects, visibility, orientation, and image-quality attributes all benefit from a common representation of the person crop, while independent heads allow each attribute to retain its own label space and decision boundary.

2.2 Deployment Setting

The intended input is a cropped person image produced by an upstream camera analytics pipeline. The model is therefore optimized for repeated inference on real camera imagery rather than isolated high-resolution photographs.

The deployment requirements shaped the choice of architecture and training strategy. In particular, the production model should:

• execute without a foundation-model dependency at inference time;
• support a conventional deployment runtime such as ONNX;
• predict all required attributes from a single shared feature extraction pass;
• remain stable under camera, illumination, scale, and image-quality variation; and
• permit the attribute heads to evolve without replacing the entire visual representation pipeline.

3. Background and Motivation

3.1 Compact Visual Backbones

ConvNeXt modernized conventional convolutional networks using design principles developed during the rise of vision transformers. ConvNeXt V2 further improved the family through representation-learning-oriented architectural and training changes. The Tiny variant provides an attractive operating point for deployment because it retains a conventional convolutional inference graph while offering substantially more representational capacity than older lightweight CNN families.

For the present task, the architecture itself is only one component. A strong generic backbone trained with insufficient domain data remains limited by the visual distribution it has seen. Our focus is therefore on how the backbone is trained rather than on increasing its size.

3.2 Contrastive Learning from Unlabeled Person Crops

SimCLR demonstrated that useful representations can be learned without class labels by encouraging two augmented views of the same image to remain close in representation space while separating representations belonging to different images.

This property is particularly useful for camera analytics. Large quantities of unlabeled person crops can be collected much more easily than multi-attribute annotations. Contrastive learning allows this data to contribute directly to the representation.

For person imagery, useful invariances include moderate changes in crop, photometric appearance, compression, scale, and other transformations that should not fundamentally alter the semantic content of the subject.

3.3 Alignment to a Stronger Self-Supervised Representation

Pure instance-level contrastive learning does not guarantee that the learned feature geometry will preserve all of the semantic distinctions useful for downstream attribute prediction.

We therefore additionally align the ConvNeXt representation with DINOv2 features. DINOv2 provides a strong self-supervised visual representation learned at a scale unavailable to the specialist model. The teacher is not used to replace the compact network. Instead, it provides a target representation during pretraining.

This produces a complementary pair of objectives:

• SimCLR encourages invariance and discrimination using the unlabeled domain corpus itself.
• DINOv2 feature alignment encourages the compact network to organize those domain images according to a richer pretrained visual representation.

At inference time, only ConvNeXt V2 Tiny remains.

3.4 Pseudo-Labels as a Supervision Source

After representation learning, the remaining problem is to learn the task-specific attribute boundaries.

A general-purpose vision-language model can produce candidate labels for many such attributes without task-specific training. However, directly treating a single generated answer as ground truth introduces substantial label noise. Small changes in prompt wording can change predictions, especially for ambiguous, low-resolution, or partially occluded samples.

We use this sensitivity as a filtering mechanism rather than ignoring it. Two separately designed prompts request the same attribute prediction. Agreement is treated as evidence that the label is stable enough to retain; disagreement causes that pseudo-label to be excluded from supervised training.

4. Awiros Person Attribute Benchmark Dataset

4.1 Dataset Scope

The released benchmark contains 35,831 person crops captured under real camera conditions.

The collection contains approximately 22.7K outdoor samples and 13.1K indoor/night samples, providing substantial coverage of different illumination and environmental regimes. The release is intentionally oriented toward real-world camera imagery, including cases that are difficult for conventional person-attribute datasets: partial subjects, non-frontal views, reduced resolution, occlusion, nighttime illumination, compression artifacts, and varying crop quality.

The benchmark contains labels for 17 person-attribute heads spanning the following broad visual categories:

• appearance;
• clothing;
• carried objects;
• accessories;
• visibility;
• orientation;
• occlusion; and
• image quality.

The purpose of the benchmark is not to measure generic image understanding. It is to provide a consistent evaluation surface for person-attribute models intended for camera analytics.

4.2 Held-Out Evaluation

The benchmark is held out from the training process used for the reported Awiros model. All systems in Section 6 are evaluated on the same benchmark samples and attribute definitions.

This common evaluation surface is particularly important when comparing specialist networks with foundation-model representations. Different pretrained models have seen different source distributions and cannot be compared meaningfully using their original reported metrics. Extracting representations from the same person crops and evaluating the same attribute targets provides a substantially more controlled comparison.

4.3 Gated Distribution

Because the dataset consists of person imagery collected under real camera conditions, it is distributed as a gated evaluation resource rather than through unrestricted public links.

Applicants are asked to identify their organization, intended use, and the system they intend to benchmark. Approved access is granted individually to the email address supplied during the request process.

The objective of gating is to make the benchmark usable by legitimate researchers and practitioners while retaining basic control over redistribution and use.

5. Training Methodology

5.1 Overview

Training is divided into two conceptually distinct stages:

1. Representation pretraining, which consumes the large unlabeled corpus and trains the ConvNeXt backbone using contrastive learning and DINOv2 alignment.
2. Attribute specialization, which generates a higher-precision pseudo-labeled subset and trains the 17 classification heads.

This separation is important. The unlabeled corpus is substantially larger than the pseudo-labeled corpus and is used to learn the general visual representation. Labels are then required only to map that representation to the attribute spaces.

5.2 SimCLR-Style Representation Learning

For a source person crop (x_i), two stochastic augmented views are generated,

$$x_i^{(1)}\sim \mathcal{T}(x_i),x_i^{(2)}\sim \mathcal{T}(x_i)\mathrm{.}x\_i^\{(1)\}\; \backslash sim\; \backslash mathcal\{T\}(x\_i),\; \backslash qquad\; x\_i^\{(2)\}\; \backslash sim\; \backslash mathcal\{T\}(x\_i).$$

Both views pass through the shared ConvNeXt encoder and a training-time projection function (q(\cdot)):

$$z_i^{(1)}=q(f_{\theta }(x_i^{(1)})),z_i^{(2)}=q(f_{\theta }(x_i^{(2)}))\mathrm{.}z\_i^\{(1)\}\; =\; q(f\_\backslash theta(x\_i^\{(1)\})),\; \backslash qquad\; z\_i^\{(2)\}\; =\; q(f\_\backslash theta(x\_i^\{(2)\})).$$

The two views of the same source crop form a positive pair, while other examples in the batch provide negatives. Using normalized representations and temperature (\tau), the contrastive objective follows the standard normalized temperature-scaled cross-entropy form:

$${\mathrm{\ell }}_{i,j}=-\log \frac{\exp (\mathrm{sim}(z_i,z_j)\mathrm{/}\tau )}{\sum _{k\mathrm{\ne }i}\exp (\mathrm{sim}(z_i,z_k)\mathrm{/}\tau )}\mathrm{.}\backslash ell\_\{i,j\}\; =\; -\backslash log\; \backslash frac\{\; \backslash exp(\backslash operatorname\{sim\}(z\_i,z\_j)/\backslash tau)\; \}\{\; \backslash sum\_\{k\backslash neq\; i\}\; \backslash exp(\backslash operatorname\{sim\}(z\_i,z\_k)/\backslash tau)\; \}.$$

The resulting loss encourages the backbone to preserve information that remains stable across augmentation while avoiding representation collapse.

The key advantage is scale: none of the samples consumed by this stage require person-attribute annotations.

5.3 DINOv2 Feature Alignment

For the same source image, a frozen DINOv2 model produces a teacher representation

$$t_i=f_{\text{DINOv2}}(x_i)\mathrm{.}t\_i\; =\; f\_\{\backslash text\{DINOv2\}\}(x\_i).$$

A projection of the ConvNeXt representation is optimized to remain close to the teacher feature. Conceptually, the alignment term may be expressed using cosine distance:

$${\mathcal{L}}_{\text{align}}=1-\cos (q_a(f_{\theta }(x_i)),t_i)\mathrm{.}\backslash mathcal\{L\}\_\{\backslash text\{align\}\}\; =\; 1\; -\; \backslash operatorname\{cos\}\; \backslash left(\; q\_a(f\_\backslash theta(x\_i)),\; t\_i\; \backslash right).$$

The complete representation-learning objective combines the two signals:

$${\mathcal{L}}_{\text{repr}}={\mathcal{L}}_{\text{SimCLR}}+{\lambda }_{\text{align}}{\mathcal{L}}_{\text{align}}\mathrm{.}\backslash mathcal\{L\}\_\{\backslash text\{repr\}\}\; =\; \backslash mathcal\{L\}\_\{\backslash text\{SimCLR\}\}\; +\; \backslash lambda\_\{\backslash text\{align\}\}\backslash mathcal\{L\}\_\{\backslash text\{align\}\}.$$

The two terms play different roles. The contrastive objective anchors learning in the actual camera-domain distribution, whereas the alignment objective transfers semantic organization from DINOv2.

Importantly, DINOv2 is required only during training. The teacher, teacher-feature extraction, and feature-alignment branch are absent from the exported production graph.

5.4 Multi-Prompt Consensus Pseudo-Labeling

Once the backbone has learned a strong domain representation, candidate supervised data is generated with a vision-language model.

For attribute (h), two prompts (p_1) and (p_2) are independently constructed to request the same label using different wording or contextual framing:

$${\tilde{y}}_{i,h}^{(1)}=V(x_i,p_1,h)\backslash tilde\{y\}^\{(1)\}\_\{i,h\}\; =\; V(x\_i,p\_1,h)$$

and

$${\tilde{y}}_{i,h}^{(2)}=V(x_i,p_2,h)\mathrm{.}\backslash tilde\{y\}^\{(2)\}\_\{i,h\}\; =\; V(x\_i,p\_2,h).$$

A pseudo-label is retained only when the predictions agree:

$$c_{i,h}=\mathbf{1}[{\tilde{y}}_{i,h}^{(1)}={\tilde{y}}_{i,h}^{(2)}]\mathrm{.}c\_\{i,h\}\; =\; \backslash mathbf\{1\}\; \backslash left[\; \backslash tilde\{y\}^\{(1)\}\_\{i,h\}\; =\; \backslash tilde\{y\}^\{(2)\}\_\{i,h\}\; \backslash right].$$

When (c_{i,h}=1), the common prediction becomes the training target. When (c_{i,h}=0), no target from that prompt pair is used for the corresponding sample and attribute.

This is deliberately a precision-oriented strategy. A single VLM pass produces more labels, but it also passes prompt-sensitive mistakes directly into the training set. Consensus filtering discards some potentially correct examples in exchange for a cleaner supervised signal.

The approach should not be interpreted as converting machine predictions into perfect ground truth. Two prompts can still share a correlated failure mode. Agreement is used as a practical confidence heuristic.

5.5 Attribute-Head Training

The consensus-filtered labels are used to train the task-specific classification heads on top of the pretrained ConvNeXt representation.

For head (h), the supervised loss is computed only for retained labels:

$${\mathcal{L}}_h=\sum _i{c}_{i,h}\mathrm{CE}(g_h(f_{\theta }(x_i)),{\tilde{y}}_{i,h})\mathrm{.}\backslash mathcal\{L\}\_\{h\}\; =\; \backslash sum\_i\; c\_\{i,h\}\; \backslash ;\; \backslash operatorname\{CE\}\; \backslash left(\; g\_h(f\_\backslash theta(x\_i)),\; \backslash tilde\{y\}\_\{i,h\}\; \backslash right).$$

The complete supervised objective aggregates the available attribute losses:

$${\mathcal{L}}_{\text{sup}}=\sum _{h=1}^{17}{w}_h{\mathcal{L}}_h\mathrm{.}\backslash mathcal\{L\}\_\{\backslash text\{sup\}\}\; =\; \backslash sum\_\{h=1\}^\{17\}\; w\_h\; \backslash mathcal\{L\}\_h.$$

This masked multi-head formulation allows a crop to contribute supervision to attributes for which consensus was obtained without requiring every generated label for the crop to be accepted.

5.6 Why the Two Stages Matter

The two-stage design avoids asking noisy pseudo-labels to solve both representation learning and task specialization simultaneously.

The first stage answers: what visual structure should the compact network preserve?

The second stage answers: where should the attribute decision boundaries lie within that representation?

Development experiments indicated that supervised-only training did not achieve the same robustness as the combined representation-pretraining pipeline. Although a complete numerical ablation table is not included in this release, this observation motivated retaining unlabeled representation learning as a first-class part of the production training recipe rather than treating it as an optional initialization step.

6. Experimental Evaluation

6.1 Evaluation Protocol

All systems are evaluated on the same held-out person-attribute benchmark.

The Awiros models operate as end-to-end multi-head classifiers: each person crop is passed through the ConvNeXt backbone and the 17 attribute heads.

The PE-Core-G14 and DINOv3 experiments use a different protocol. Their visual encoders remain frozen. Features are extracted from each model and an XGBoost classifier is trained for each attribute. For DINOv3 ViT-7B/16, both a global-average-pooled representation and the CLS-token representation are evaluated separately.

These results therefore measure the usefulness of the frozen representations under a common downstream probe. They are not equivalent to end-to-end fine-tuning of the corresponding foundation models.

For attribute head (h), top-1 accuracy is

$$A_h=\frac{1}{N_h}\sum _{i=1}^{N_h}\mathbf{1}[{\widehat{y}}_{i,h}=y_{i,h}]\mathrm{.}A\_h\; =\; \backslash frac\{1\}\{N\_h\}\; \backslash sum\_\{i=1\}^\{N\_h\}\; \backslash mathbf\{1\}\; [\backslash hat\{y\}\_\{i,h\}=y\_\{i,h\}].$$

We report two aggregate metrics:

$$A_{\text{mean}}=\frac{1}{17}\sum _{h=1}^{17}{A}_h,A\_\{\backslash text\{mean\}\}\; =\; \backslash frac\{1\}\{17\}\; \backslash sum\_\{h=1\}^\{17\}\; A\_h,$$

and the median attribute accuracy

$$A_{\text{median}}=\mathrm{median}(A_1,\dots ,A_{17})\mathrm{.}A\_\{\backslash text\{median\}\}\; =\; \backslash operatorname\{median\}\; (A\_1,\backslash ldots,A\_\{17\}).$$

Reporting both reduces the risk of describing a 17-task system using a single statistic that may be disproportionately affected by a small number of easy or difficult heads.

6.2 Compared Systems

The comparison contains five configurations:

Awiros ConvNeXt V2 Tiny (Current).
The current production-oriented multi-head model trained using the representation-pretraining and consensus-supervision pipeline described in Section 5.

PE-Core-G14.
A frozen representation model evaluated by fitting XGBoost attribute classifiers to extracted features.

Awiros ConvNeXt (Previous).
The previous Awiros person-attribute model, included to quantify progress within the production model lineage.

DINOv3 ViT-7B/16 — GAP.
Frozen DINOv3 representations with global average pooling followed by XGBoost attribute classifiers.

DINOv3 ViT-7B/16 — CLS.
The same frozen model using its CLS representation for the downstream probes.

6.3 Benchmark Results

System	Evaluation Setup	Mean Accuracy	Median Accuracy
Awiros ConvNeXt V2 Tiny (Current)	End-to-end multi-head classifier	82.18%	82.75%
PE-Core-G14	XGBoost classifiers on frozen representations	81.52%	82.62%
Awiros ConvNeXt (Previous)	End-to-end multi-head classifier	80.70%	81.96%
DINOv3 ViT-7B/16 (GAP)	XGBoost classifiers on frozen representations	78.75%	80.24%
DINOv3 ViT-7B/16 (CLS)	XGBoost classifiers on frozen representations	75.72%	78.49%

The current Awiros model obtains the best aggregate result in the comparison, reaching 82.18% mean accuracy and 82.75% median accuracy.

Relative to PE-Core-G14, the current model improves mean accuracy by 0.66 percentage points and median accuracy by 0.13 percentage points.

Relative to the previous Awiros ConvNeXt model, the new model improves mean accuracy by 1.48 percentage points and median accuracy by 0.79 percentage points.

The margin over DINOv3 is larger under the evaluated frozen-feature protocol. Compared with global-average-pooled DINOv3 ViT-7B/16 features, the current Awiros model is 3.43 percentage points higher in mean accuracy. Compared with the CLS-token probe, the difference is 6.46 percentage points.

Per-Attribute Accuracy Comparison

The following table reports top-1 accuracy for each attribute on the same held-out benchmark. All values are percentages; bold indicates the highest score in each row.

Attribute	Awiros ConvNeXt (Previous)	Awiros ConvNeXt V2 Tiny (Current)	PE-Core-G14	DINOv3 ViT-7B/16 (CLS)	DINOv3 ViT-7B/16 (GAP)
Gender	89.99%	91.57%	96.00%	91.47%	93.89%
Age	71.91%	73.50%	82.62%	70.73%	70.85%
Top color	72.35%	74.76%	70.77%	43.10%	45.54%
Bottom color	74.25%	75.19%	60.80%	50.12%	58.65%
Hair length	84.08%	85.52%	87.07%	81.91%	84.20%
Sleeve length	84.49%	86.15%	84.35%	81.57%	84.78%
Bottom length	88.39%	89.00%	86.04%	82.19%	86.76%
Has backpack	93.44%	94.23%	95.04%	92.44%	79.05%
Has handbag	90.50%	91.34%	93.70%	90.27%	95.21%
Head accessory	81.96%	84.66%	78.53%	78.49%	93.07%
Face accessory	88.70%	88.79%	87.93%	87.94%	87.93%
Visibility	81.53%	81.42%	80.75%	78.12%	80.24%
Orientation	80.98%	82.32%	77.95%	75.15%	78.65%
Is occluded	82.58%	82.75%	83.75%	80.98%	82.92%
Image quality	80.18%	77.44%	79.62%	78.17%	79.19%
Top attire	58.34%	64.33%	70.66%	62.71%	67.43%
Bottom attire	68.29%	74.07%	70.38%	61.94%	70.48%

6.4 Interpreting the Foundation-Model Comparison

The most important caveat in Table 1 is also what makes the result useful.

The foundation models are not fine-tuned end-to-end. Their representations are frozen and a downstream XGBoost classifier is learned for each attribute. The Awiros model, by contrast, is a dedicated end-to-end attribute classifier.

The result should therefore not be read as evidence that ConvNeXt V2 Tiny is universally stronger than PE-Core-G14 or DINOv3. A sufficiently resourced end-to-end adaptation of those architectures represents a different experiment.

The result supports a narrower and more practically relevant conclusion: for this real-world person-attribute benchmark, the compact specialist model provides attribute information at least as effectively as the evaluated frozen foundation-model representations under the common probe protocol.

That operating point is valuable because the large representation models are not required in the production inference path.

7. Deployment

7.1 Production Lineage

Awiros person-attribute models from this development lineage have been deployed in production since mid-2024.

The current ConvNeXt V2 Tiny model is the latest model in that lineage. Production history matters for this task because camera analytics models encounter forms of variation that are difficult to capture completely in a static benchmark: new camera sensors, changing scene layouts, nighttime transitions, compression changes, partial crops, unusual clothing, and evolving operating environments.

The released benchmark is intended to make a significant subset of that real-world difficulty measurable and reproducible.

7.2 Inference Architecture

The exported production model contains the compact ConvNeXt V2 backbone and its attribute heads. Neither the DINOv2 teacher nor the vision-language model used during data creation is required during inference.

This distinction allows heavyweight models to contribute during training while preserving a conventional specialist deployment graph:

$$\text{Person crop}\to \text{ConvNeXt V2 Tiny}\to \text{17 attribute predictions}\mathrm{.}\backslash text\{Person\; crop\}\; \backslash rightarrow\; \backslash text\{ConvNeXt\; V2\; Tiny\}\; \backslash rightarrow\; \backslash text\{17\; attribute\; predictions\}.$$

The model is distributed in ONNX format, making it suitable for integration with production inference runtimes without requiring the original training framework.

7.3 Training-Time Scale, Inference-Time Compactness

The training pipeline deliberately separates the compute budget used to build the model from the compute budget required to serve it.

Large models contribute in two offline roles:

1. DINOv2 supplies representation targets during unlabeled pretraining.
2. A vision-language model supplies candidate attribute labels during supervised-data generation.

Once training is complete, both are removed.

This is the key production trade-off: expensive representation models can be used to improve the specialist offline, while the deployed system retains the latency, memory, and operational characteristics of a compact ConvNeXt classifier.

8. Model and Benchmark Release

The release consists of two related artifacts.

8.1 Awiros Person Attribute Recognition Model

• Architecture: ConvNeXt V2 Tiny
• Task: Multi-head person attribute classification
• Number of attribute heads: 17
• Deployment format: ONNX
• Pipeline category: Image classification
• Primary domain: Person crops from real-world CCTV and camera systems
• Training strategy: SimCLR-style unlabeled pretraining, DINOv2 representation alignment, and consensus-filtered pseudo-label supervision

8.2 Awiros Person Attribute Benchmark Dataset

• Benchmark size: 35,831 person crops
• Environment coverage: Outdoor plus indoor/night imagery
• Annotations: 17 person-attribute targets
• Distribution: Gated
• Purpose: Evaluation and comparison of person-attribute recognition systems under real camera conditions

8.3 Access

The model and benchmark are distributed through controlled access.

Request model and benchmark access

Applicants are asked to provide their name, organization, intended use, and the model or system they plan to evaluate.

Approved users are granted access individually through a restricted Google Drive or Zoho WorkDrive folder using the email address supplied in the request. Files are not distributed through unrestricted public-link sharing.

9. Limitations and Responsible Use

Several limitations should be considered when interpreting the reported results.

First, the benchmark is intentionally biased toward the operating conditions of real camera systems. This makes it useful for deployment-oriented evaluation, but the results should not be assumed to transfer directly to studio photography, web imagery, or other visual domains without validation.

Second, mean and median top-1 accuracy summarize performance across 17 heads but do not expose every failure mode. Attribute-level results, per-class precision and recall, class imbalance, and calibration can provide additional information and should be considered when evaluating a model for a particular application.

Third, multi-prompt agreement reduces pseudo-label noise but does not eliminate it. The two prompts are evaluated by the same underlying class of vision-language system and may share correlated visual or semantic errors. Consensus is therefore a filtering heuristic rather than a guarantee of correctness.

Fourth, the foundation-model comparisons in this report are representation-probe experiments. They should not be extrapolated into a claim about the performance of end-to-end fine-tuned PE-Core or DINOv3 systems.

Fifth, a person-attribute model can produce confident errors under severe occlusion, very low resolution, unusual clothing, or distribution shifts. Attribute predictions should therefore be treated as probabilistic visual observations rather than infallible facts about a person.

Finally, the benchmark and model involve visual analysis of people. The release is intended for legitimate computer-vision research, benchmarking, and responsible video-analytics development. Attribute predictions should not be used as the sole basis for consequential decisions about individuals, and the system is not designed for determining personal identity.

10. Conclusion

This work releases a real-world benchmark and a compact production-oriented model for person attribute recognition.

The benchmark contains 35,831 person crops spanning 17 attribute heads and difficult camera conditions including outdoor, indoor, night, viewpoint, visibility, occlusion, and image-quality variation.

The model is built around a simple principle: use labels only where labels are necessary.

A substantially larger corpus of unlabeled person imagery is first used to learn the visual representation through a SimCLR-style contrastive objective and alignment with DINOv2. Supervision is introduced only afterward, using a vision-language model to generate candidate labels and a two-prompt consensus rule to suppress unstable pseudo-labels. The resulting labels train the task-specific attribute heads.

Under the released evaluation protocol, the current ConvNeXt V2 Tiny model reaches 82.18% mean and 82.75% median top-1 accuracy across 17 attributes. It outperforms the evaluated PE-Core-G14 and DINOv3 frozen-representation probes in aggregate while maintaining a compact end-to-end inference architecture.

The broader takeaway is not that specialist networks make foundation models unnecessary. In this pipeline, foundation-scale representations are valuable precisely because they can be used during model construction. The deployment model benefits from their supervision without inheriting their inference footprint.

For production computer vision, this separation is useful: large models can be teachers, label generators, and representation sources; the deployed model does not have to be large itself.

Awiros person-attribute models from this lineage have operated in production since mid-2024. The current release makes both the resulting model and a common real-world evaluation surface available for external benchmarking.

References

[1] S. Woo et al., “ConvNeXt V2: Co-designing and Scaling ConvNets with Masked Autoencoders,” Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023.

[2] T. Chen, S. Kornblith, M. Norouzi, and G. Hinton, “A Simple Framework for Contrastive Learning of Visual Representations,” Proceedings of the International Conference on Machine Learning (ICML), 2020.

[3] M. Oquab et al., “DINOv2: Learning Robust Visual Features without Supervision,” arXiv:2304.07193, 2023.

[4] D.-H. Lee, “Pseudo-Label: The Simple and Efficient Semi-Supervised Learning Method for Deep Neural Networks,” ICML Workshop on Challenges in Representation Learning, 2013.

[5] T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting System,” Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016.