Adaptive Multi-Granularity Mixture Consistency with Dynamic Prototype-Guided Rebalancing for Long-Tailed Visual Recognition
Keywords:
Long-tail recognition, Imbalanced data, Image classification, deep learning, Features, SABLC, Multi-granularity Augmentation, Consistency Learning, Prototype Rebalancing,Abstract
Long-tailed visual recognition remains challenging i.e., severe classfrequency imbalance pushes deep models toward head-class accuracy
at the expense of tail classes. Recent one-stage methods improve this
along two separate fronts i.e., mixture-consistency augmentation, which
pairs global and local image views for more robust representations, and
adaptive rebalancing, which adjusts class weights using training-derived
signals rather than fixed schedules. Each, however, has an open limitation: mixture-based methods such as GLMC use a small, fixed set
of augmentation scales regardless of an image’s semantic complexity,
while existing dynamic rebalancing schemes rely on hand-tuned epoch
schedules rather than a direct, representation-level measure of how
well each class is currently learned. We present ADMG-LTR (Adaptive
Multi-Granularity Mixture Consistency with Dynamic Prototype-Guided
Rebalancing), a one-stage framework combining and extending both
directions. An Adaptive Multi-Granularity Mixture (AMGM) module
replaces the fixed global/local view pairing with a learnable scale selection, where a lightweight gating network sets each view’s spatial extent
from a per-image complexity estimate. A Consistency Alignment Loss
(CAL) extends pairwise global-local consistency to enforce agreement
jointly across all scale. K granularity levels. Dynamic Prototype-Guided
Rebalancing (DPGR) maintains per-class momentum prototypes and uses
prototype-to-sample distance as a live signal to modulate class weights,
replacing epoch-scheduled or frequency-based weighting with one that
tracks each class’s evolving representation quality directly. We evaluate
ADMG-LTR on standard long-tailed benchmarks across multiple imbalance ratios against representative one-stage baselines, including GLMC
(mixture consistency) and AREA (adaptive reweighting), to assess the
effectiveness of combining learnable multi-granularity augmentation
