How Reliable Is Automatic Emotion Classification in Children’s Drawings? A Reproducible Benchmark on a Public Corpus with Calibration and Selective Prediction

Abstract

Emotion recognition in children’s drawings is difficult because affect is carried by sparse strokes, symbolic objects, and overall composition rather than by the stable appearance statistics of photographs. We built a reproducible four-class benchmark (Angry, Fear, Happy, Sad) on a single public corpus of 818 children’s drawings and compared three transfer-learning regimes under identical stratified five-fold splits with nested model selection: ResNet-50, ViT-B/16, and an end-to-end fine-tuned SigLIP image encoder (SigLIP-FT). SigLIP-FT reached the highest macro-F1 (0.773 ± 0.028), ahead of ViT-B/16 (0.700 ± 0.040) and ResNet-50 (0.598 ± 0.038), and was the best calibrated (ECE 0.119). Frozen-feature linear probes preserve this ordering (0.541, 0.648, 0.731), locating the advantage in the pretrained representations rather than in fine-tuning, while zero-shot SigLIP reaches only 0.510, so task-specific supervision remains necessary. Margin-based abstention raised retained-set macro-F1 to 0.810 at 78.0% coverage and 0.844 at 61.4% coverage—post hoc operating points computed on the pooled out-of-fold predictions; deployment thresholds must be fixed on independent data—and SigLIP-FT attains the lowest area under the risk–coverage curve (AURC 0.126 versus 0.199 and 0.278). Residual errors are highly structured: 87.0% lie within the negative-emotion triad, and Fear is the hardest category for all three architectures. All findings are established on this single corpus; their transfer to other collections is an open question. Coverage–performance behavior, rather than a single full-coverage score, is the appropriate reporting standard for ambiguous visual domains of this kind.