Abstract
Most existing person re-identification (re-id) methods rely on supervised model learning on per-camera-pair manually labelled pairwise training data. This leads to poor scalability in a practical re-id deployment, due to the lack of exhaustive identity labelling of positive and negative image pairs for every camera-pair. In this work, we present an unsupervised re-id deep learning approach. It is capable of incrementally discovering and exploiting the underlying re-id discriminative information from automatically generated person tracklet data end-to-end. We formulate an Unsupervised Tracklet Association Learning (UTAL) framework. This is by jointly learning within-camera tracklet discrimination and cross-camera tracklet association in order to maximise the discovery of tracklet identity matching both within and across camera views. Extensive experiments demonstrate the superiority of the proposed model over the state-of-the-art unsupervised learning and domain adaptation person re-id methods on eight benchmarking datasets.
Introduction
Person Re-Identification (re-id) is the critical task of matching individuals across non-overlapping camera views. Traditionally, this relies on supervised learning, which requires exhaustive manual labeling of image pairs across every camera combination. This approach faces a massive scalability bottleneck: as camera networks expand, the labeling effort grows quadratically. While domain adaptation offers a partial solution by transferring knowledge from labeled datasets, it assumes environmental similarity that often does not exist in the real world. This research addresses the "pure" unsupervised challenge: enabling a model to recognize people across a network using only raw video data without any human intervention.
The UTAL Framework
The proposed Unsupervised Tracklet Association Learning (UTAL) framework operates on tracklets—short sequences of a person's movement captured by a single camera's automated tracker. The framework integrates two primary components to learn discriminative features without identity labels.
Per-Camera Tracklet Discrimination (PCTD)
UTAL treats each camera as a separate task in a multi-task learning setup. Initially, every tracklet is assigned a unique ID. However, because automated trackers often fail (fragmentation), the same person may be split into multiple tracklets. To mitigate this, UTAL employs "Soft Labeling." Instead of rigid IDs, it assigns probability distributions based on visual similarity. If two tracklets within the same camera view are visually similar, the model learns to treat them as the same identity.
Cross-Camera Tracklet Association (CCTA)
This global step searches for reciprocal nearest neighbors across different cameras. When Tracklet A in Camera 1 and Tracklet B in Camera 2 identify each other as their best match, the model assumes they represent the same person. Their feature representations are then pulled closer together in the embedding space. Training follows a curriculum learning strategy, establishing local features (PCTD) before introducing complex cross-camera associations (CCTA) to avoid convergence on false matches.
Experimental Evaluation
The researchers evaluated UTAL across eight major benchmarks, including video-based datasets (iLIDS-VID, PRID2011, MARS) and image-based datasets (CUHK03, Market-1501, DukeMTMC-ReID, MSMT17). A new benchmark, DukeTracklet, was introduced to test the model against raw, automatically generated tracklets containing real-world tracking errors and occlusions.
Key Results
UTAL significantly outperformed existing unsupervised and domain-adaptation methods:
- Market-1501: Achieved 69.2% Rank-1 accuracy, surpassing the previous best unsupervised method by 7%.
- MARS: On this large-scale video dataset, it reached 49.9% Rank-1 accuracy, a 3.1% improvement over the previous state-of-the-art.
- Robustness: The model demonstrated high effectiveness on the DukeTracklet dataset, proving its ability to handle imperfect data from real-world automated trackers.
Discussion and Future Work
UTAL overcomes the "Catch-22" of unsupervised re-id—needing good features to find matches while needing matches to learn good features—by starting with high-confidence local matches and progressing to harder cross-camera ones. The multi-task setup implicitly forces the model to learn view-invariant features even before explicit cross-camera matching begins.
Despite these gains, a performance gap remains between unsupervised and fully supervised models (which can exceed 90% accuracy). The method's reliance on initial tracking quality suggests that extremely crowded scenes might introduce too much noise for the soft labels to converge effectively. Future research may incorporate spatio-temporal constraints, such as travel time between cameras, to further refine the unsupervised association process.
Summary updated