Two videos. Same bug?
The answer is in the details.
Different bugs can look nearly alike.
And true duplicates can take entirely different steps to get there.
In crowd testing, 80%+ are duplicates.
Prior research found most crowd-sourced bug reports repeat a known bug.
vs
Read the whole scene.
A Vision Transformer, trained self-supervised with DINO and fine-tuned on 66k app screens, attends to GUI components and how they’re arranged.
ViT + DINObag of visual wordsTF-IDF
Find the text. Read it.
EAST localizes the text on every frame; TrOCR, a Transformer OCR model, reads it. Lucene scores the overlap.
EASTTrOCRLucene
Align. Weigh the ending.
A weighted longest-common-substring match lines the videos up end to end and counts later frames more, since that’s where bugs show.
REPORT A · 8 FRAMES
REPORT B · 6 FRAMES
earlier framesframe weight →later frames
Evaluated on nine Android apps
0
duplicate-detection tasks
0
videos watched, on average, to find the first duplicate
89.8%mean reciprocal rank
+8.7%mRR over TANGO, statistically significant
8 of 9apps improved