VTF900
A large-scale benchmark dataset for visible-thermal infrared face recognition, containing paired imagery of 900 subjects. It supports 1:1 verification and 1:N closed-set recognition evaluation and is used for cross-spectral and cross-dataset analysis.
Timeline (last 90 days)
VTF900 is a cross-spectral face recognition benchmark dataset containing paired visible-thermal infrared imagery of 900 subjects, supporting large-scale evaluation of both the verification task (1:1 matching) and the closed-set identification task (1:N matching).
Experiments on VTF900 adopt a cross-dataset protocol of training on the ARL-VTF dataset and evaluating on VTF900.
At the full VTF900 gallery scale (N=900) under frontal conditions, the ViT-B model achieves 97.68% AUC and 51.10% TAR@1%FAR.
At the full VTF900 gallery scale (N=900) under non-frontal poses, DPIT-ResNet50 achieves 94.68% AUC and 36.08% TAR@1%FAR.
The VTF900 study shows that high AUC under small-scale galleries does not translate into reliable performance at deployment scale, and model rankings derived from hundred-person-scale galleries may undergo substantial reordering at the 900-person scale.
VTF900 covers benchmark analysis of three architecture categories: convolutional networks (baselines), vision Transformers, and domain-adaptive neural networks.
The VTF900 dataset includes expression, pose, and occlusion variations, among which pose and occlusion are the most common sources of failure for cross-spectral recognition in real-world scenarios.
All Facts (7)
VTF900 is a cross-spectral face recognition benchmark dataset containing paired visible-thermal infrared imagery of 900 subjects, supporting large-scale evaluation of both the verification task (1:1 matching) and the closed-set identification task (1:N matching).
50%UnverifiedExperiments on VTF900 adopt a cross-dataset protocol of training on the ARL-VTF dataset and evaluating on VTF900.
50%UnverifiedAt the full VTF900 gallery scale (N=900) under frontal conditions, the ViT-B model achieves 97.68% AUC and 51.10% TAR@1%FAR.
50%UnverifiedAt the full VTF900 gallery scale (N=900) under non-frontal poses, DPIT-ResNet50 achieves 94.68% AUC and 36.08% TAR@1%FAR.
50%UnverifiedThe VTF900 study shows that high AUC under small-scale galleries does not translate into reliable performance at deployment scale, and model rankings derived from hundred-person-scale galleries may undergo substantial reordering at the 900-person scale.
50%UnverifiedVTF900 covers benchmark analysis of three architecture categories: convolutional networks (baselines), vision Transformers, and domain-adaptive neural networks.
50%UnverifiedThe VTF900 dataset includes expression, pose, and occlusion variations, among which pose and occlusion are the most common sources of failure for cross-spectral recognition in real-world scenarios.
50%