MCCL: Multiscale Contrastive Learning for DFER

• The paper introduces MCCL, a supervised contrastive learning method that integrates multiscale feature aggregation, category-aware weighting, and dynamic temperature to mitigate long-tailed bias in dynamic facial expression recognition.
• MCCL constructs instance-level representations through scale-specific pooling and projection, achieving notable improvements in UAR and feature discrimination on benchmarks like DFEW.
• The method emphasizes balancing head and tail classes within the MICACL framework without relying on memory banks or explicit cross-scale pairing, thereby enhancing overall system robustness.

Searching arXiv for the cited MCCL-related papers to ground the article in the current literature. Multiscale Category-aware Contrastive Learning (MCCL) is a supervised contrastive learning strategy introduced within the MICACL framework for long-tailed dynamic facial expression recognition (DFER). In that formulation, MCCL combines multiscale feature construction, category-aware positive weighting, and a dynamic temperature to balance optimization between major and minor categories, with the stated aim of reducing induction bias caused by class imbalance and improving robustness and generalization on in-the-wild DFER benchmarks (Cui et al., 4 Sep 2025). The acronym is not unique across the literature: closely related or homonymous usages appear in semantic segmentation, anomaly detection, video-based generalized category discovery, and open-vocabulary segmentation, so the exact meaning of MCCL is context-dependent (Pissas et al., 2022).

1. Definition and terminological scope

In MICACL, MCCL is explicitly named Multiscale Category-aware Contrastive Learning and is described as a supervised contrastive learning strategy tailored to long-tailed DFER. Its defining elements are: category-aware weights that reflect both class imbalance and per-sample difficulty, a dynamic temperature that reduces head-class dominance in similarity computation, and multiscale aggregation of contrastive signals constructed from the instance-level representation produced by GEIIM and WIAN (Cui et al., 4 Sep 2025).

The acronym is reused elsewhere with different expansions or related meanings. This is not merely terminological variation; it changes the role of supervision, the type of multiscale structure, and the operational meaning of “category-aware.”

Paper	Official term	Domain
(Cui et al., 4 Sep 2025)	Multiscale Category-aware Contrastive Learning	Long-tailed DFER
(Jing et al., 8 Sep 2025)	Memory-guided Consistency-aware Contrastive Learning	Video-GCD
(Zhang et al., 8 Aug 2025)	Multi-Category Contrastive Learning	Open-vocabulary semantic segmentation
(Fan et al., 2024)	Local and Global Class-aware Contrastive Learning	Multi-class anomaly detection

A related line of work in semantic segmentation applies supervised contrastive learning on encoder features at multiple stages and introduces cross-scale local-global constraints, but the paper title does not use the acronym MCCL; rather, the details characterize it as a multiscale, category-aware supervised contrastive formulation over segmentation features (Pissas et al., 2022). By contrast, the graph paper “Multi-Scale Subgraph Contrastive Learning” does not include category-aware components in the provided document and explicitly lacks labels, pseudo-labels, prototypes, or class constraints (Liu et al., 2024).

2. Problem setting and motivation

The MICACL formulation of MCCL is motivated by the long-tailed nature of DFER. The paper states that head classes have many videos while tail classes have few samples, producing severe class imbalance and model induction bias, with decision boundaries skewed toward head classes and poor feature discrimination for tails (Cui et al., 4 Sep 2025). Existing methods are described as improving DFER performance while often failing to address long-tailed category distributions explicitly.

The paper positions MCCL against three families of imbalance handling strategies. Class-balanced loss and LDAM reweight samples or margins at the classifier or cross-entropy level. Focal loss downweights easy examples uniformly. Supervised contrastive learning (SupCon) improves representation learning by bringing same-class features together, but in standard form it implicitly favors head classes because those classes provide many positives. MCCL is presented as operating directly in representation space with dynamic, category-aware contrastive weighting and temperature, thereby targeting imbalance-induced bias during feature shaping for both head and tail classes (Cui et al., 4 Sep 2025).

This design implies that MCCL is not simply a multiscale feature pyramid with a contrastive head. Its distinguishing premise is that the same supervised contrastive objective should not contribute equally across categories and samples when the label distribution is long-tailed. A plausible implication is that the method’s “category-awareness” is concentrated in the weighting and temperature mechanisms rather than in explicit prototype banks or cross-scale category graphs.

3. Formal construction

MCCL in MICACL is defined on the instance-level feature matrix produced after multi-instance interaction and aggregation:

$\mathbf{X}' \in \mathbb{R}^{B \times C},$

where $B$ is the mini-batch size and $C$ is the feature dimension (Cui et al., 4 Sep 2025).

The category-aware weight combines inverse class frequency and per-sample difficulty:

$$\mathbf{W}_{\mathrm{c}(i)} = \frac{1}{\mathrm{n}_{c}[i]} \cdot \Big(1 - \mathrm{softmax}(\mathbf{y}_{\mathrm{p}})[i]\Big).$$

Here, $\mathrm{n}_{c}[i]$ denotes the number of samples in category $y_i$ in the training set, and $\mathrm{softmax}(\mathbf{y}_{\mathrm{p}})[i]$ is the predicted probability for the ground-truth class of sample $i$. The paper states that this increases the contribution of tail categories through $1/n_c$ and hard examples through $1-$ predicted probability (Cui et al., 4 Sep 2025).

Multiscale features are constructed by scale-specific pooling and projection:

$B$0

with the scale-$B$1 embedding for sample $B$2 denoted by $B$3 (Cui et al., 4 Sep 2025). The pairwise similarity at scale $B$4 is cosine similarity with temperature $B$5:

$B$6

The temperature is dynamically adjusted using class-frequency statistics:

$B$7

The paper states that this reduces the influence of head classes on similarity computation and emphasizes tail-category features (Cui et al., 4 Sep 2025).

Positive pairs are supervised and defined within a mini-batch:

$B$8

For each scale $B$9, the supervised contrastive loss is

$C$0

The multiscale loss averages over scales:

$C$1

MCCL is then augmented with a Classification Enhancement Term (CET):

$C$2

In the broader training setup, the paper uses cross-entropy classification as the baseline task loss and gives the joint objective as

$C$3

An important technical clarification is that the multiscale behavior comes from aggregating scale-specific losses. The paper explicitly notes that cross-scale effects are realized by aggregating the scale-specific losses; no explicit cross-scale pairs are formed (Cui et al., 4 Sep 2025). This distinguishes the method from cross-scale local-global contrastive formulations in semantic segmentation, where anchors at one resolution are contrasted directly with features from another resolution (Pissas et al., 2022).

4. Position within MICACL

MICACL consists of three parts: Graph-Enhanced Instance Interaction Module (GEIIM), Weighted Instance Aggregation Network (WIAN), and MCCL (Cui et al., 4 Sep 2025). The input video is sampled to 16 frames, and a 3D backbone (R3D with Torchvision pretraining) extracts per-frame or per-clip features.

GEIIM models inter-instance spatio-temporal interaction through a dynamic adjacency matrix:

$C$4

where $C$5 and $C$6 (Cui et al., 4 Sep 2025). WIAN then aggregates instance features with dynamic weights derived from GEIIM outputs:

$C$7

followed by LSTM cell updates

$C$8

A multi-head self-attention module then produces the global instance-level representation $C$9, which becomes the input to MCCL (Cui et al., 4 Sep 2025).

The training workflow reported in the paper is explicit. It uses AdamW, a cosine scheduler, 300 epochs, initial LR $$\mathbf{W}_{\mathrm{c}(i)} = \frac{1}{\mathrm{n}_{c}[i]} \cdot \Big(1 - \mathrm{softmax}(\mathbf{y}_{\mathrm{p}})[i]\Big).$$0, min LR $$\mathbf{W}_{\mathrm{c}(i)} = \frac{1}{\mathrm{n}_{c}[i]} \cdot \Big(1 - \mathrm{softmax}(\mathbf{y}_{\mathrm{p}})[i]\Big).$$1, and weight decay 0.05 (Cui et al., 4 Sep 2025). The paper also states that no memory bank/queue is described and that long-tailed balancing is handled by $$\mathbf{W}_{\mathrm{c}(i)} = \frac{1}{\mathrm{n}_{c}[i]} \cdot \Big(1 - \mathrm{softmax}(\mathbf{y}_{\mathrm{p}})[i]\Big).$$2 and $$\mathbf{W}_{\mathrm{c}(i)} = \frac{1}{\mathrm{n}_{c}[i]} \cdot \Big(1 - \mathrm{softmax}(\mathbf{y}_{\mathrm{p}})[i]\Big).$$3 inside MCCL rather than via a sampler.

This placement inside MICACL matters methodologically. GEIIM and WIAN shape a global instance representation through adaptive adjacency matrices, multiscale convolutions, weighted aggregation, and temporal modeling; MCCL then regularizes that representation according to label-aware contrastive structure. This suggests that MCCL is best understood as the long-tail optimization component of MICACL rather than as a standalone feature extractor.

5. Empirical evidence

The paper reports ablation results on DFEW showing the marginal contribution of MCCL. The Baseline (no GEIIM/WIAN/MCCL) achieves WAR 64.55, UAR 54.17. GEIIM+WIAN (no MCCL) reaches WAR 68.62, UAR 56.26. GEIIM+WIAN+MCCL (full MICACL) reaches WAR 69.91, UAR 64.34. The paper explicitly states that MCCL yields a notable gain in UAR (+8.08 points over GEIIM+WIAN), which it interprets as better balance across classes, especially tail classes (Cui et al., 4 Sep 2025).

For overall benchmarks, the paper reports the following results. On DFEW, MICACL achieves WAR 69.56 and UAR 63.21, described as state-of-the-art UAR. On FERV39k, MICACL reaches WAR 48.57 and UAR 40.25 (Cui et al., 4 Sep 2025). The paper further highlights tail-category improvement on DFEW, noting Disgust (from prior 0.00 in several baselines to 29.65) and Fear (up to 42.57).

The reported visualization evidence is consistent with the contrastive objective: t-SNE visualization (Fig. 3) shows tighter intra-class clusters and clearer inter-class boundaries with MICACL, which the paper associates with MCCL’s role in shaping more balanced, discriminable features (Cui et al., 4 Sep 2025). In computational terms, the paper reports overall FLOPs for MICACL of 1.69G, described as comparable to M3DFEL (1.65G).

A plausible implication is that MCCL’s strongest measurable effect is on UAR, not merely on overall accuracy. That pattern is consistent with the method’s stated goal of balancing head and tail categories rather than only improving dominant-class recognition.

6. Related formulations and neighboring research

Several related formulations clarify what MCCL is and is not. In semantic segmentation, “Multi-scale and Cross-scale Contrastive Learning for Semantic Segmentation” applies supervised contrastive learning over multiple encoder stages, maps them to a common embedding space, and introduces a cross-scale supervised local-global loss linking high-resolution local features to lower-resolution global features (Pissas et al., 2022). That method is explicitly category-aware because positives are defined by class labels at each scale, and it operates without memory banks or heavy data augmentation. However, its contrastive structure differs from MICACL’s MCCL because it uses cross-scale pair construction rather than scale-wise averaging alone.

In multi-class anomaly detection, “Revitalizing Reconstruction Models for Multi-class Anomaly Detection via Class-Aware Contrastive Learning” describes MCCL as equivalent to Local and Global Class-aware Contrastive Learning (LGC): local patch-level contrastive learning across multiple encoder feature pyramid levels and global instance-level contrastive learning in a bottleneck representation space (Fan et al., 2024). There, “multiscale” refers to feature hierarchy, and “category-aware” is provided by raw object category labels such as carpet or wood.

The graph paper “Multi-Scale Subgraph Contrastive Learning” is relevant for multiscale structure but not for category-awareness. The provided summary explicitly states that the document does not include any explicit category-aware components and that category-aware contrastive learning is not explicitly present (Liu et al., 2024). By contrast, the video-based generalized category discovery paper uses the acronym MCCL for Memory-guided Consistency-aware Contrastive Learning, where multiscale structure arises from vertical clusterings at $$\mathbf{W}_{\mathrm{c}(i)} = \frac{1}{\mathrm{n}_{c}[i]} \cdot \Big(1 - \mathrm{softmax}(\mathbf{y}_{\mathrm{p}})[i]\Big).$$4 and category-aware behavior from per-class prototypes and logits memory (Jing et al., 8 Sep 2025). SynSeg uses MCCL to mean Multi-Category Contrastive Learning, and its own summary explicitly states that it is multi-category, not multiscale (Zhang et al., 8 Aug 2025).

These neighboring formulations show that “MCCL” has become a family resemblance term rather than a single canonical method. Across domains, the recurring motifs are supervised or pseudo-supervised pair construction, category-conditioned feature separation, and either multiscale feature hierarchies or multi-granular relational structure. The precise mechanism, however, varies substantially by task.

7. Limitations, misconceptions, and future directions

A common misconception is to treat MCCL as a generic label-aware contrastive loss that automatically implies prototype learning, cross-scale pair construction, or memory queues. In MICACL, none of those are defining properties. The paper states that no memory bank/queue is described, that scale interactions are obtained by averaging scale-specific losses, and that balancing is achieved through $$\mathbf{W}_{\mathrm{c}(i)} = \frac{1}{\mathrm{n}_{c}[i]} \cdot \Big(1 - \mathrm{softmax}(\mathbf{y}_{\mathrm{p}})[i]\Big).$$5 and dynamic $$\mathbf{W}_{\mathrm{c}(i)} = \frac{1}{\mathrm{n}_{c}[i]} \cdot \Big(1 - \mathrm{softmax}(\mathbf{y}_{\mathrm{p}})[i]\Big).$$6 rather than through effective-number formulas or LDAM margins (Cui et al., 4 Sep 2025).

The paper also records explicit limitations. It notes future work to further optimize spatio-temporal modeling and explore advanced long-tail strategies. It further states that MCCL’s dynamic temperature and category-aware weighting are heuristic and do not yet incorporate, for example, effective-number formulas or memory-augmented contrastive queues, and that the CET details are relegated to the Appendix (Cui et al., 4 Sep 2025). These remarks delimit the current method’s scope: it is a practically specified optimization component with strong benchmark behavior, but not a complete theory of long-tail contrastive calibration.

More broadly, the literature suggests two axes along which the idea may evolve. One axis is cross-scale structure, as in segmentation methods that contrast high-resolution local features with low-resolution global features (Pissas et al., 2022). The other is memory or prototype guidance, as in video-GCD and anomaly detection formulations that use feature prototypes, logits prototypes, or global class context (Jing et al., 8 Sep 2025). This suggests that future versions of MCCL in the MICACL sense may combine inverse-frequency weighting and dynamic temperature with more explicit class memory, adaptive scale weighting, or prototype-level supervision.