Assessment of State-of-the-Art Deep Neural Networks on a Multi-Era, Cross-Domain Dataset of Indian Cinema

Authors

  • Khushi Modi Author
  • Devang Pandya Author
  • Rachna Modi Author
  • Ketan Sarvakar Author

Keywords:

Facial Expression Recognition, Affective Computing, Cross Domain Generalization, Indian Cinema, Bidirectional GRU, Cross Stream Attention, Dual-Stream Architecture

Abstract

This study sets out to close two long-standing gaps in video-based Facial Expression Recognition (FER): the field's reliance on Western subjects, and its silence on the visual evolution of non-Western cinema over time. The standard benchmarks (FER2013, AffectNet, IEMOCAP, and MELD) were all built from Western subjects, and none of them transfer cleanly to Indian cinematic material. Indian screen performance is organized around the classical Navarasa system rather than the Ekman categories (Ekman & Friesen, 1971) underlying these benchmarks, and Hindi cinema spans a 52-year visual transition from monochrome to RGB that no existing dataset reflects.

Addressing both gaps directly, this study introduces three original Indian cinematic FER datasets, namely Ittefaq (1969, black-and-white), 420 IPC (2021, HD colour), and Taarak Mehta Ka Ooltah Chashmah (Television), built through a reproducible six-stage pipeline, and proposes HybridMicroDCSA, a family of dual-stream networks that fuse spatial and motion-difference streams through Cross-Stream Attention (CSA).

Four model families, namely HybridMicroDCSA, MobileNetV3, EfficientNet-B0, and ResNet-50, are evaluated across a 15 scenario framework covering single-domain, cross-domain, multi-domain, and replay-based training. On the 1969 black-and-white domain, the 2.15M-parameter MicroDCSA v4 reaches 74.44% accuracy, beating ResNet-50 (71.11%, 25.6M parameters) at 12.45x parameter efficiency, and wins 12 of 15 scenarios overall.

The 52-year domain shift causes a 27.51% accuracy drop under direct transfer, of which a cross-domain replay strategy recovers 14.05 points, reaching a 69.60% cross-domain average. Because these figures come from test sets as small as 15 samples per class and have not been subjected to formal significance testing, they should be read as an initial benchmark rather than a definitive ranking.

Downloads

Published

2026-07-07