The presence of clinically significant clutter is a core diagnostic indicator of hoarding disorder. Ideally, the assessment of clutter in a dwelling requires direct visual inspection or reliance on photographs to provide ratings using a validated measure, which is labor-intensive, subjective, and only approximately repeatable. A Clutter Image Rating (CIR) scale was proposed to improve clutter-assessment consistency, but still requires ”manual” comparison of a living space to a set of reference images. Automatic clutter classification from images would be a useful decision-support proxy, yet the task is difficult because labeled data are limited, labels are ordinal, and neighboring CIR levels can differ only by subtle changes in object density, obstruction, and stacking. The most successful clutter-assessment method to date is based on the Vision Transformer (ViT). In this work, we improve on the RGB-only ViT baseline through partial transformer updating, tuned optimization, and geometry-preserving augmentation, while keeping the benchmark’s loss function and evaluation protocol unchanged. On the HINDER-2025 clutter-image dataset, this tuning improves performance by 6.28 percentage points in classification accuracy, while an additional DINOv2 initialization further boosts the accuracy by 1.01 percentage points. We also investigate whether clutter prediction can be improved by fusing auxiliary cues derived from the same RGB image, without requiring new manual annotations to enlarge the dataset. We evaluate the impact of structural cues from object detection, semantic cues from clutter-focused image masking, and volumetric cues from relative depth estimation. The structural and semantic cues provide modest gains over the tuned RGB baseline, while the volumetric cue provides the largest improvement. With DINOv2 initialization, the model using RGB+Depth reaches 63.42% accuracy, improving over the RGB-only ViT baseline by 9.96 percentage points. Impressively, this model achieves 96.27% accuracy within ±1 from the ground truth, which is important in practice since professionals acknowledge challenges in assigning exact CIR values. Effectively, this level of accuracy suggests that an automated algorithm can serve as a reliable proxy for assessment by professionals, thus potentially leading to reduced labor costs, elimination of subjectivity and precise repeatability.