Random Tracklet Sampling in Gait Recognition

• Random Tracklet Sampling is a temporal strategy that balances local motion continuity and global sequence diversity for robust gait recognition.
• It utilizes a discrete distribution over tracklet lengths, assembling multiple short tracklets to capture varied walking states in challenging conditions.
• Empirical results demonstrate that RTS improves rank-1 accuracy by up to 1.1% over conventional methods in wild surveillance datasets.

Searching arXiv for the named papers to ground the article and citations. Random Tracklet Sampling (RTS) is a temporal sampling strategy for gait recognition that was introduced in "TrackletGait: A Robust Framework for Gait Recognition in the Wild" (Zhang et al., 4 Aug 2025). It is designed for silhouette sequences that are non-periodic, noisy, and occluded, conditions under which conventional gait-cycle assumptions are often invalid. RTS samples multiple short tracklets from random positions in a sequence, thereby combining short-term temporal continuity with broader temporal coverage. In the cited literature, the term also has adjacent but distinct meanings: in multiple object tracking, random sampling of candidate tracklets is used for inpainting missing trajectory segments rather than for gait-sequence construction, while some tracklet-centric video systems do not use random tracklet sampling at all (Saleh et al., 2020).

1. Problem setting and motivation

RTS is motivated by the mismatch between conventional gait sampling assumptions and real surveillance footage. Earlier gait pipelines often assume a stable gait cycle and rely on sampling consecutive frames around that cycle. The underlying premise is that the walking sequence is periodic and sufficiently clean that a contiguous temporal window captures identity-related motion (Zhang et al., 4 Aug 2025).

TrackletGait explicitly argues that this premise fails in the wild. Real sequences may contain stopping, turning, and non-walking motion, viewpoint changes, scale changes, severe occlusion, background subtraction errors, and non-periodic segments within the same sequence. Under such conditions, a cycle-aligned or purely local sampler can overconcentrate on a short temporal state, miss important walking variations, or emphasize unrepresentative frames. RTS is proposed as a response to this setting-specific failure mode rather than as a generic replacement for all temporal samplers (Zhang et al., 4 Aug 2025).

The design objective is stated as a balance between robustness and representation. A short tracklet preserves local temporal information, while sampling several such tracklets from different positions increases coverage of diverse walking states. This suggests that RTS is best understood as a compromise between local motion fidelity and global sequence diversity rather than as a purely stochastic augmentation heuristic.

2. Formal definition and sampling procedure

TrackletGait represents a gait silhouette sequence as

$X = \{x_t\}_{t=1}^{M},$

where $M$ is the number of frames and $x_t$ is the silhouette frame at time $t$. A mini-batch samples $N$ frames from $X$, producing $\mathcal{X}$ (Zhang et al., 4 Aug 2025).

The RTS procedure begins by choosing a tracklet length $l$ from a discrete distribution $P(l)$. For the configuration reported in the paper when $N=32$,

$M$0

with

$M$1

A single sampled tracklet is then defined as

$M$2

where $M$3 is a random starting index, $M$4 is the tracklet length, and $M$5 is the step size between frames. The paper describes the default sampling as consecutive short clips and notes that RTS generalizes prior methods using $M$6 (Zhang et al., 4 Aug 2025).

To reach the total of $M$7 sampled frames, the method draws $M$8 tracklets such that

$M$9

and combines them as

$x_t$0

The full RTS output is

$x_t$1

Operationally, each training iteration may therefore use a different tracklet length, while every sampled input is assembled from multiple short temporal segments drawn from different positions in the original sequence (Zhang et al., 4 Aug 2025).

3. Relation to consecutive and random-frame sampling

RTS is presented as a generalization of two prior strategies: Consecutive Sampling (CS) and Random Frame Sampling (RFS) (Zhang et al., 4 Aug 2025).

CS samples one contiguous block,

$x_t$2

where $x_t$3 is randomly chosen from $x_t$4. Its advantage is local continuity, but in non-periodic sequences it may cover only one motion state and ignore others.

RFS samples independent frames,

$x_t$5

where each $x_t$6 is sampled independently from $x_t$7. Its advantage is diversity, but it discards temporal continuity entirely and can be overly sensitive to unreliable silhouettes (Zhang et al., 4 Aug 2025).

RTS interpolates between these extremes. If

$x_t$8

RTS reduces to CS. If

$x_t$9

RTS reduces to RFS. This formalizes the claim that RTS balances robustness and representation: it retains temporal context within each short tracklet while preventing the entire sample from collapsing to one contiguous span or to isolated frames. The paper’s stated intuition is that “A short tracklet contains more than three frames to preserve local temporal information. Furthermore, the shorter the length of a sequence, the more tracklets can be sampled (with a fixed total sampling frames), allowing us to cover a greater variety of walking states.” (Zhang et al., 4 Aug 2025)

A common misconception is to treat RTS as merely “random clipping.” The formal reduction above shows that the method is instead a parameterized family whose limiting cases recover existing samplers.

4. Dataset-dependent behavior and empirical findings

TrackletGait reports a direct ablation on Gait3D. Starting from a baseline with no RTS, no HWD, and no HE triplet, the reported rank-1 accuracy is $t$0. Adding RTS only yields $t$1; adding RTS and HWD yields $t$2; adding RTS, HWD, and HE Triplet yields $t$3 (Zhang et al., 4 Aug 2025). The difference between the baseline and the RTS-only model is $t$4, which the paper uses to isolate the contribution of the temporal sampling redesign.

The paper also analyzes fixed-length variants. On Gait3D, RTS-8 is best; on GREW, RTS-32 is best; and on OU-MVLP, CASIA-B, CCPG, and SUSTech1K, RTS-16 tends to be best (Zhang et al., 4 Aug 2025). The discussion connects these outcomes to dataset complexity and temporal stability. Gait3D is described as having highly interrupted walking, so shorter tracklets and more random positions help capture diverse states. GREW is described as containing pedestrians moving toward or away from the camera, so longer tracklets help capture gradual appearance change. The more stable datasets favor intermediate lengths.

To support this interpretation, the paper introduces AIFV. It defines

$t$5

and then

$t$6

The paper states that lower AIFV means more stable periodicity, whereas higher AIFV means more variation and less regularity (Zhang et al., 4 Aug 2025). This suggests that tracklet length should not be regarded as universally optimal; rather, it interacts with the degree of interruption, scale variation, and temporal regularity in the target dataset.

5. Role within the TrackletGait framework

RTS is one of the three core components of TrackletGait, alongside Haar Wavelet-based Downsampling and Hardness Exclusion Triplet Loss (Zhang et al., 4 Aug 2025). Within that framework, RTS addresses the temporal dimension specifically: it changes how the model sees a sequence during training, without being described as a post hoc retrieval mechanism or a trajectory-repair stage.

The practical configuration reported for RTS uses $t$7 sampled frames, the distribution $t$8, $t$9, and $N$0, and silhouettes resized to $N$1. The paper describes this as a “balanced parameter configuration across different scenarios” (Zhang et al., 4 Aug 2025). The final reported results of TrackletGait-64 are $N$2 rank-1 on Gait3D and $N$3 rank-1 on GREW, with only $N$4M backbone parameters.

Within the paper’s framing, RTS contributes by improving robustness to wild surveillance conditions and by increasing coverage of diverse walking states. A plausible implication is that the method is particularly valuable when sequence quality is heterogeneous within a single identity instance, because its multi-tracklet construction prevents the input from being dominated by one local failure mode.

6. Related uses and terminological boundaries

The phrase “random tracklet sampling” is not uniform across tracklet-based vision research. In "Probabilistic Tracklet Scoring and Inpainting for Multiple Object Tracking" (Saleh et al., 2020), the relevant operation is not gait-sequence sampling but autoregressive generation of candidate missing trajectory segments. ArTIST models each tracklet as an autoregressive discrete distribution over motion velocities and, during inpainting, samples $N$5 plausible tracklets from a multinomial distribution over clustered motion classes. The appendix states that multinomial sampling uses $N$6 candidate tracklets, and the sampled continuations are filtered by a tracklet rejection scheme based on IoU (Saleh et al., 2020). In that setting, random sampling is a mechanism for long-occlusion recovery and identity preservation, not a temporal sampler for constructing training inputs.

By contrast, "ChatVideo: A Tracklet-centric Multimodal and Versatile Video Understanding System" does not describe any random tracklet sampling scheme (Wang et al., 2023). The system detects all tracklets exhaustively, stores them in a database, and uses a deterministic representative-frame selection rule for appearance annotation:

$N$7

For motion annotation, it splits each tracklet into equal-length temporal segments, stated in implementation to be 32 frames. The paper explicitly distinguishes random ordering of database records in a prompt from any random selection of tracklets from the video (Wang et al., 2023).

These contrasts clarify a recurring source of ambiguity. In gait recognition, RTS denotes a training-time temporal sampling strategy over silhouette sequences. In multiple object tracking, random sampling of tracklets may refer instead to stochastic generation of candidate trajectory bridges across occlusions. In tracklet-centric video understanding, tracklets may be exhaustive database units with deterministic frame selection and no random tracklet sampling at all.