Coarse Attribute Prediction with Task Agnostic Distillation for Real World Clothes Changing ReID

TL;DR

This work tackles CC-ReID under real-world low-quality artifacts by introducing RLQ, a framework that combines Coarse Attribute Prediction (CAP) and Task Agnostic Distillation (TAD). CAP learns coarse pose and gender cues from RGB inputs to enrich representations without adding test-time noise, while TAD uses an external HQ dataset to distill robust HQ-like features into the target model, mitigating LQ feature clustering. The approach yields consistent Top-1 improvements across PRCC, LTCC, LaST, and DeepChange benchmarks, with notable gains on PRCC and competitive results on LTCC, supported by extensive ablations. Overall, RLQ offers a practical strategy to improve real-world CC-ReID robustness by leveraging coarse attributes and HQ-LQ cross-domain supervision during training.

Abstract

This work focuses on Clothes Changing Re-IDentification (CC-ReID) for the real world. Existing works perform well with high-quality (HQ) images, but struggle with low-quality (LQ) where we can have artifacts like pixelation, out-of-focus blur, and motion blur. These artifacts introduce noise to not only external biometric attributes (e.g. pose, body shape, etc.) but also corrupt the model's internal feature representation. Models usually cluster LQ image features together, making it difficult to distinguish between them, leading to incorrect matches. We propose a novel framework Robustness against Low-Quality (RLQ) to improve CC-ReID model on real-world data. RLQ relies on Coarse Attributes Prediction (CAP) and Task Agnostic Distillation (TAD) operating in alternate steps in a novel training mechanism. CAP enriches the model with external fine-grained attributes via coarse predictions, thereby reducing the effect of noisy inputs. On the other hand, TAD enhances the model's internal feature representation by bridging the gap between HQ and LQ features, via an external dataset through task-agnostic self-supervision and distillation. RLQ outperforms the existing approaches by 1.6%-2.9% Top-1 on real-world datasets like LaST, and DeepChange, while showing consistent improvement of 5.3%-6% Top-1 on PRCC with competitive performance on LTCC. *The code will be made public soon.*

Figures (18)

• Figure 1: Noisy fine-grained attributes: Samples from LaST, LTCC, DeepChange, PRCC dataset.
• Figure 2: Robustness against Low-Quality (RLQ) (a) Fine-grained attribute (e.g. pose) is discretized/clustered into coarse classes which the model predicts using RGB images for learning attribute feature representation. Classification is an easy task that avoids input noise from attributes. (b) External (${\bigcirc}$) HQ embeddings pull LQ ones towards it via self-supervision & distillation, dissolving the LQ cluster on the target dataset ($\triangle$, $\square$), especially helpful when the target dataset doesn't have HQ-LQ pairs.
• Figure 3: ReID against Pose and Gender: LTCC mis-matches (in red) for CAL gu2022clothes.
• Figure 4: Robustness against Low-Quality (RLQ): Base ReID model is our baseline with RGB-only input (black arrows). The base model is trained alternately between: Upper: Coarse Attributes Prediction (CAP) trained on target ReID dataset with train-only pose branch for pose disentanglement (pose clusters generated via AlphaPose (brown)) and gender classifier (gender assigned manually) for gender entanglement. Bottom: Task Agnostic Distillation (TAD) trains student Base model using teacher Base model (pink, distillation) and self-supervision on classification logits (green). Both CAP and TAD (White modules) are train-only, with Base Model $\boldsymbol{f_{ReID}}$ used during RGB-only inference. The colored arrow indicates input.
• Figure 5: Pose Clustering: left: Pose Vector (AlphaPose fang2017rmpe) consists of the length and angle of 13 body lines, which are used in clustering (K-means). right: 2 pose clusters shown.