OU-ISIR Large-Population Gait Dataset

• OU-ISIR Large-Population Dataset is a comprehensive gait database featuring variants designed for age-related analysis and cross-view recognition.
• It includes clearly defined subsets like OULP-C1V1 and OULP-C1V2, providing thousands of gait cycles and multi-angle silhouette sequences for robust evaluations.
• The dataset has enabled significant insights into gait recognition and age estimation, with strong correlations observed between gait metrics and chronological age.

Searching arXiv for the cited OU-ISIR-related papers to ground the article in the current literature. arxiv.search({"query":"id:(Velankar, 15 Jul 2025) OR id:(Chen et al., 2017) OR OU-ISIR Large Population dataset", "max_results": 10}) The OU-ISIR Large-Population Dataset is a gait database available through Osaka University’s biometric database and cited as a public benchmark for gait-age studies and cross-view gait recognition. In the literature summarized here, two closely related forms are emphasized: the OU-ISIR Gait Database, Large Population with Age, used for large-scale correlation analysis between chronological age and interpretable gait metrics, and the OU-ISIR large population gait dataset in the OULP-C1V1 and OULP-C1V2 versions, used for silhouette-sequence-based cross-view recognition (Velankar, 15 Jul 2025, Chen et al., 2017).

1. Dataset identity and reported variants

The dataset is described in one line of work as the OU-ISIR Gait Database, Large Population with Age, and in another as the OU-ISIR large population gait dataset (Velankar, 15 Jul 2025, Chen et al., 2017). The latter is described as the largest gait dataset suitable for training deep neural networks, while the former is used as a large-scale public benchmark to examine how gait changes with age.

The following variants and population figures are explicitly reported.

Variant	Reported size	Reported note
OULP-C1V1	4,007 subjects	used for all experiments in the cross-view recognition study
OULP-C1V2	4,016 subjects	more accurate bounding box and different moving-average filter size in size-normalized silhouette creation
Large Population with Age	63,846 gait cycles / samples analyzed	used for correlation analysis; not necessarily 63,846 unique subjects

For the age-oriented benchmark, the paper does not provide a detailed sensor specification in the body text beyond its being a gait database used for feature correlation analysis. It further states that the cited OU-ISIR Large-Population with Age dataset is typically a video/silhouette-based gait dataset used in computer-vision gait research. In that study, it is used as a source of gait measurements from which spatiotemporal and pose-derived features can be derived (Velankar, 15 Jul 2025).

Data availability is reported through the DOI https://doi.org/10.1000/oulp-age.

2. Experimental configuration in cross-view recognition

The cross-view recognition study evaluates on OULP-C1V1 and follows the protocol of Mansur et al. (Chen et al., 2017). It uses a subset of OULP-C1V1 containing 1912 subjects. Each subject has sequences from four view angles: 55°, 65°, 75°, and 85°.

For each subject in that subset, there are probe and gallery gait sequences. Each subset has four sequences corresponding to the four view angles, so each subject has 8 total sequences: 4 probe-view sequences and 4 gallery-view sequences. The sequence length ranges from 19 to 43 frames.

The reported train/test organization is as follows:

• 856 subjects for training
• 956 subjects for testing/evaluation
• 100 subjects are randomly selected as a validation set

The source also notes an internal inconsistency in the arithmetic around the validation split. The defensible statement is therefore the one explicitly reported: 856 training subjects, 100 validation subjects chosen from training, and 956 test subjects.

Cross-view comparisons are performed over all combinations of probe and gallery views, that is, $\{55, 65, 75, 85\} \times \{55, 65, 75, 85\}$, yielding 16 cross-view recognition tasks. This multi-view structure makes the dataset suitable for evaluating both identification and verification under view mismatch, especially in large view-gap cases such as 55° versus 85° (Chen et al., 2017).

3. Use in gait-age analysis

In gait-based age estimation, the OU-ISIR Large-Population with Age dataset is used as the main public dataset for a large-scale correlation analysis between chronological age and gait features. Its purpose is not to train a classifier directly, but to quantify how strongly age is associated with interpretable gait metrics and to support the claim that age-related changes are visible in gait patterns (Velankar, 15 Jul 2025).

The study reports analysis of 63,846 gait cycles / samples from the OU-ISIR Large-Population dataset. A wording inconsistency is also noted: the future-work section incorrectly says “train on the large OU-ISIR set (63 846 subjects),” which is likely a misuse of “subjects” for “samples/cycles.” The most consistent interpretation is therefore 63,846 gait cycles/samples were analyzed, not necessarily 63,846 unique subjects.

Five key gait metrics are reported:

1. Stride length
2. Walking speed / gait speed
3. Step cadence
4. Step-time variability, referred to in the paper as step time variance
5. Joint-angle entropy

These metrics are described as classic gait descriptors known to change with aging. The reported Pearson correlations are:

Feature	Pearson $r$	p-value
Stride length (m)	$-0.45$	$< 0.0001$
Cadence (steps/min)	$-0.38$	$< 0.0001$
Step time variance	$+0.31$	$< 0.0001$
Joint angle entropy	$+0.27$	$< 0.0001$
Gait speed (m/s)	$r$0	$r$1

The reported Pearson correlation coefficient is

$r$2

where $r$3 is age for sample $r$4, $r$5 is the gait feature value for sample $r$6, and $r$7, $r$8 are sample means. In this context, $r$9 means the feature decreases as age increases, and $-0.45$0 means the feature increases as age increases.

All five relationships are reported with

$-0.45$1

indicating that these relationships are highly unlikely to be due to chance. The paper interprets the negative correlations for stride length, gait speed, and cadence as indicating that gait becomes slower and more cautious with age, while the positive correlations for step-time variability and joint-angle entropy indicate that gait becomes less regular and more variable, consistent with reduced motor consistency and increased movement complexity or instability (Velankar, 15 Jul 2025).

4. Silhouette-sequence recognition methodology on OULP-C1V1

The cross-view recognition work processes raw silhouette sequences directly, rather than collapsing a sequence into a GEI. The method is a CNN + feature-map pooling + Siamese-style matching architecture (Chen et al., 2017).

For a pair of gait sequences, the reported pipeline is:

1. Each silhouette frame is resized and fed through a CNN feature extractor (fCNN).
2. The difference image between adjacent silhouettes is also fed to the same fCNN.
3. Per-frame feature maps are aggregated across time using feature map pooling.
4. The two sequence-level fused feature maps are compared by absolute difference.
5. A second CNN (mCNN) maps this difference to a compact similarity feature.
6. A fully connected classifier outputs whether the two sequences belong to the same person.

The per-frame feature extractor (fCNN) has 2 convolutional layers. The first has 16 filters with filter size $-0.45$2 and stride 1. The second has 64 filters with filter size $-0.45$3 and stride 1. After each convolution there is $-0.45$4 spatial pooling and local response normalization (LRN). The output after the second conv + pooling stage is 64 feature maps of size $-0.45$5.

For a sequence $-0.45$6, the feature tensor for frame $-0.45$7 is written as

$-0.45$8

with output size

$-0.45$9

The temporal aggregation mechanism is feature map pooling. The max-pooling form is

$< 0.0001$0

and the mean-pooling alternative is

$< 0.0001$1

The reported empirical result is that mean pooling performs about 5% worse than max pooling, so max pooling is used in all experiments.

After pooling, the two fused feature maps are compared by element-wise absolute difference and passed through mCNN, which has 1 convolutional layer, 256 filters, and filter size $< 0.0001$2. Its output size is

$< 0.0001$3

which is reshaped into a vector with 112896 elements. A fully connected layer then predicts the final two-class output. The loss is reported as the negative log-likelihood

$< 0.0001$4

where $< 0.0001$5 are target indicators and $< 0.0001$6 are the predicted probabilities for the two classes: non-matching pair and matching pair (Chen et al., 2017).

5. Quantitative performance and benchmark behavior

On OULP-C1V1, silhouettes are resized to $< 0.0001$7, no data augmentation is used, the mini-batch size is 128, the learning rate is 0.001, momentum is 0.0, optimization is SGD, training length is up to 1.8 million iterations, training time is 7 days, the implementation is in Torch 7, and training is performed on NVIDIA GeForce GTX Titan X. Validation is performed every 100 iterations, and the best model is selected by highest recognition rate (Chen et al., 2017).

The method is designed for arbitrary-length sequences, and the study examines sequence lengths from 1 frame to 43 frames. It reports that longer sequences generally improve performance.

For the 956-subject test set, pairwise matching between probe and gallery sequences requires about

$< 0.0001$8

pairwise measurements for each task. To reduce runtime, fused features are precomputed and stored, and similarities are computed only between fused features.

The reported evaluation metrics are Rank-1, Rank-2, Rank-5 identification rates and Equal Error Rate (EER). Mean results by probe view are:

Probe view	Mean Rank-1	Mean EER
55°	83.1%	2.27%
65°	93.0%	1.41%
75°	90.5%	1.39%
85°	83.0%	1.70%

Additional reported observations are specific. The hardest cross-view case is typically 55° probe to 85° gallery, with Rank-1 = 62.2% and EER = 3.74%. Best EERs are around 1.15%. Same-view or closer-view combinations are much higher, often above 94% in Rank-1.

The comparison against LDA, DATER, GMLDA, MvDA, and CCA shows that the proposed method is significantly better in EER than all compared methods. At the same time, the paper notes that in identification rate, some view-aware methods such as GMLDA and MvDA can outperform it slightly because they use view information. Despite being blind to view angle, the method is reported to achieve strong Rank-1 performance, and the paper claims approximately two times better verification accuracy than other methods (Chen et al., 2017).

6. Interpretation, limitations, and recurrent caveats

The OU-ISIR Large-Population with Age analysis is presented as evidence that age estimation from gait is not just a black-box classification problem. The dataset supports the empirical claim that aging affects spatiotemporal gait parameters, motion regularity, and joint kinematics, and that these effects are detectable at scale in public data (Velankar, 15 Jul 2025).

The age-estimation work further reports that a fine-tuned ResNet-34 with Grad-CAM attends to the knees and pelvic region. The authors interpret this as evidence that the CNN is attending to regions known to exhibit age-related gait changes, providing qualitative support for the OU-ISIR correlation findings. The saliency maps are used to argue that the model is not relying on spurious background cues, but on biomechanically meaningful body regions. A plausible implication is that the dataset is valuable not only for predictive benchmarking but also for relating learned representations to interpretable gait structure.

Several caveats recur across the reported uses of OU-ISIR. Many studies reuse public datasets like OU-ISIR, and the meta-analysis notes that results are weighted to avoid double-counting the same subjects across studies. The ambiguity between subjects and samples in the age-analysis paper can be misleading. The OU-ISIR results show significant associations, but correlation does not imply prediction performance on that dataset. More generally, the review emphasizes differences between lab and field data, even though OU-ISIR itself is treated as a benchmark dataset. The literature review also acknowledges some subjectivity in paper selection (Velankar, 15 Jul 2025).

In the cross-view recognition setting, the method evaluated on OULP-C1V1 is not explicitly view-invariant, pairwise evaluation across many subjects is computationally heavy without feature caching, and the absence of data augmentation may limit generalization. These caveats do not negate the benchmark role of OU-ISIR; rather, they delimit what can be inferred from strong performance on a large-population, multi-view, silhouette-sequence dataset (Chen et al., 2017).