[Paper Review] Re-ID done right: towards good practices for person re-identification

Jon Almazán, Bojana Gajić+2 more|arXiv (Cornell University)|Jan 16, 2018

Video Surveillance and Tracking MethodsComputer Science31 references93 citations

TL;DR

The paper identifies and validates a set of practical design and training principles for learning a global image representation for person re-identification, achieving state-of-the-art results without explicit alignment or attention modules.

ABSTRACT

Training a deep architecture using a ranking loss has become standard for the person re-identification task. Increasingly, these deep architectures include additional components that leverage part detections, attribute predictions, pose estimators and other auxiliary information, in order to more effectively localize and align discriminative image regions. In this paper we adopt a different approach and carefully design each component of a simple deep architecture and, critically, the strategy for training it effectively for person re-identification. We extensively evaluate each design choice, leading to a list of good practices for person re-identification. By following these practices, our approach outperforms the state of the art, including more complex methods with auxiliary components, by large margins on four benchmark datasets. We also provide a qualitative analysis of our trained representation which indicates that, while compact, it is able to capture information from localized and discriminative regions, in a manner akin to an implicit attention mechanism.

Motivation & Objective

• Identify a set of design and training practices that improve re-ID performance when building a global image representation.
• Evaluate the impact of architecture choices, data augmentation, and training strategy on re-ID accuracy.
• Show that a simple, well-trained global descriptor can outperform complex methods on multiple benchmarks.
• Provide qualitative insights into what the learned embedding attends to in clothing and other discriminative cues.

Proposed method

• Use a three-stream Siamese network with shared weights to produce embeddings from input triplets (query, positive, negative).
• Train with a ranking triplet loss to push positives closer to the query than negatives by a margin.
• Upscale input images and avoid distortion; accumulate gradients to maintain effective batch size when processing variable-sized images.
• Pre-train backbones on ImageNet and further fine-tune for identity classification before the ranking objective (curriculum learning).
• Employ cut-out data augmentation to improve robustness to occlusion and regularization; apply hard triplet mining to select challenging training samples.
• Demonstrate key design choices through ablative studies and compare against state-of-the-art methods across four datasets.

Experimental results

Research questions

• RQ1 Can a simple global representation, when trained with carefully designed practices, outperform more complex re-ID models?
• RQ2 Which architectural and training choices most significantly impact re-ID performance?
• RQ3 Does curriculum learning and hard triplet mining contribute substantially to convergence and accuracy in re-ID?
• RQ4 To what extent can a high-resolution, non-aligned global embedding capture discriminative cues for robust re-ID across datasets?

Key findings

• Achieve state-of-the-art results on four benchmarks; for Market-1501, mAP 81.2% (vs. 72.9–73.1 by many baselines) and up to 8.1 percentage points improvements over prior art.
• On Market MQ, mAP 92.2% and rank-1 94.7%; on Duke-reID, mAP 72.8% and rank-1 87.3% (significant gains over prior art).
• On the Person Search dataset, mAP 92.6% with a strong rank-1 of 79.7% (14.7 points above previous best).
• ResNet-101/152 backbones provide noticeable gains over ResNet-50; pretraining for identity classification before ranking substantially improves results.
• Cut-out data augmentation is crucial and outperforms standard flips/crops; large, undistorted input resolutions (416 px) are important for best performance.
• Implicit attention emerges from the learned embedding, localizing discriminative regions (e.g., clothing edges, backpacks) without explicit alignment or attention modules.