Kus669/ViSceT5-ViTextVQA
Visual Question Answering • 0.3B • Updated • 52
ViSceT5 on ViTextVQA: final model and module ablations.
Note Final model: QA-ViT + Consformer + DOVE + OACT (ocr_both). Test EM 36.01 / F1 58.93
Note + Consformer. Test EM 32.55
Note + QA-ViT. Test EM 32.88
Note + DOVE (batch 32x1). Test EM 34.60
Note QA-ViT + Consformer + DOVE, no OACT. Test EM 34.85
Note OACT scenario 1 (ocr_cross)