Technical Summary: Contrastive Order Learning (ConOrd)

Problem Statement

Ordinal regression aims to estimate the discrete or continuous rank of an object instance, a task fundamental to applications such as facial age estimation, blind image/video quality assessment, and health status scoring. Unlike standard classification, ordinal regression requires modeling the inherent order and relative distances between labels. However, existing methods face significant limitations:

• Order Learning: Techniques like Geometric Order Learning (GOL) explicitly model label ordinality but often rely on local, margin-based pairwise comparisons. These approaches typically generate no gradient once a margin is satisfied, failing to exploit the richer global ordinal context available at the batch level.
• Contrastive Learning: While supervised contrastive learning effectively leverages all samples in a batch, it typically treats labels as categorical, ignoring the inherent ordering. Extensions like Rank-N-Contrast (RnC) introduce rank-aware pairs but rely on hard thresholds, collapsing ordinal distances into binary decisions (positive/negative) and treating samples with moderate and substantial rank gaps equally as negatives.

Consequently, there is a need for a framework that captures fine-grained ordinal relationships across all sample pairs while maintaining global consistency in the embedding space.

Methodology: Contrastive Order Learning (ConOrd)

The authors propose ConOrd, a framework that integrates the strengths of order learning and contrastive learning. The core innovation is a contrastive order loss ($L_{ConOrd}$) that assigns soft, rank-aware weights to all sample pairs within a batch, rather than relying on hard positive/negative assignments.

Key Components:

1. Embedding Space with Reference Points: The framework employs an encoder $h$ mapping instances to feature vectors on a unit sphere. It introduces $M$ learnable reference points $\{\mu_m\}$, where each $\mu_m$ represents a prototype for rank $m$.
2. Soft Affinity and Disparity Weights: Unlike RnC, which uses hard thresholds, ConOrd defines continuous weights based on the rank difference $|r_i - r_j|$:
   • Affinity Weight ($a_{ij}$): Decreases as the rank gap increases, promoting attraction between samples with similar ranks. It is instantiated as $a_{ij} = \frac{1}{(r_i - r_j)^2 + \epsilon}$.
   • Disparity Weight ($b_{ij}$): Increases as the rank gap increases, promoting repulsion between samples with distant ranks. It is instantiated as $b_{ij} = (r_i - r_j)^2$.
3. The Loss Function: The loss is formulated as:
   $$L_{ConOrd} = -\frac{1}{2N} \sum_{i=1}^{2N} \frac{1}{|A(i)|} \log \frac{\sum_{j \in A(i)} a_{ij} e^{\kappa_{ij}/\tau}}{\sum_{j \in A(i)} b_{ij} e^{\kappa_{ij}/\tau}}$$
   where $\kappa_{ij}$ is the similarity (negative squared Euclidean distance) and $\tau$ is a temperature parameter. This formulation ensures that:
   • Samples with small rank gaps are attracted.
   • Samples with large rank gaps are repelled.
   • The magnitude of the rank gap modulates the strength of attraction or repulsion, enabling fine-grained ordinal supervision.
4. Regularization: A center loss ($L_{center}$) is added to encourage compact clusters around the reference points $\mu_m$ for each rank.
5. Inference: Rank estimation is performed using a $k$-nearest neighbor ($k$-NN) rule in the embedding space, averaging the ranks of the $k$ nearest training neighbors.

Key Contributions

1. Framework Integration: Introduction of ConOrd, a general framework that unifies order learning and contrastive learning within a single batch-wise formulation.
2. Novel Loss Design: Proposal of a contrastive order loss that utilizes soft affinity and disparity weights based on rank differences. This allows all sample pairs in a batch to contribute to learning, modulated by their ordinal relationships, avoiding the information loss inherent in hard thresholding.
3. State-of-the-Art Performance: Demonstration that ConOrd achieves state-of-the-art results across diverse ordinal regression benchmarks, including facial age estimation, blind image quality assessment (BIQA), and blind video quality assessment (BVQA).

Experimental Results

The authors evaluated ConOrd on multiple datasets and tasks:

• Facial Age Estimation: Tested on six datasets (MORPH II, CLAP2015, AgeDB-DIR, UTK, CACD, Adience). ConOrd achieved the best Mean Absolute Error (MAE) on all datasets except CACD, outperforming strong baselines like OrdinalCLIP and NumCLIP (which use textual guidance) and SLACE (which uses class-distribution information).
• Blind Image Quality Assessment (BIQA): Evaluated on five datasets (BID, CLIVE, KonIQ10k, SPAQ, FLIVE). ConOrd outperformed existing methods in both Spearman's Rank Order Correlation Coefficient (SRCC) and Pearson's Linear Correlation Coefficient (PCC) on all datasets, surpassing complex transformer-based and Mamba-based architectures.
• Blind Video Quality Assessment (BVQA): Evaluated on five benchmarks (LSVQ, KoNViD-1k, LIVE-VQC, CVD2014, YouTube-UGC). ConOrd consistently achieved the best SRCC and PCC scores, outperforming methods that utilize multiple backbone networks (e.g., ModularBVQA) while using fewer backbones.
• Ablation Studies:
  • Comparisons with $L_{SupCon}$, $L_{RnC}$, and other contrastive losses confirmed that the soft-weighted, all-pair comparison is superior.
  • The combination of $L_{ConOrd}$ and $L_{center}$ yielded the best results, with $L_{center}$ providing intra-class compactness.
  • The proposed quadratic weighting functions were found to be robust and effective compared to alternative functional forms.
  • The method demonstrated resilience to reduced training data and data corruption.

Significance and Claims

The paper claims that ConOrd provides a flexible foundation for ordinal representation learning by bridging the gap between the global consistency of contrastive learning and the explicit ordinal modeling of order learning.

• Generalization: The framework is presented as a general solution applicable to various ordinal regression scenarios, not limited to a specific domain.
• Efficiency: The method is computationally efficient, avoiding the need for data augmentation and dynamic negative sampling required by some baselines, leading to faster training times.
• Stability: The soft-weighting mechanism is claimed to provide stable optimization, avoiding the gradient instability associated with naive logarithmic losses or the information loss of hard thresholds.

The authors conclude that by incorporating rank-aware attraction and repulsion through soft weights, ConOrd enables all sample pairs to contribute to learning in a principled manner, resulting in robust and high-performing ordinal regression models.