Bug reports are becoming videos.
So are the duplicates.
Is this one new?
Text-based duplicate detectors can’t read a screen recording.
By hand: about 20 seconds a video.
In the paper’s user study, people watched 4.3 videos on average before finding a duplicate.
Introducing
TANGO
It takes two: what a video shows, and what it says.
01 · WHAT IT SHOWS
Learn what app screens look like.
SimCLR, trained on 66k+ RICO screenshots, maps every frame to a visual word. Each video becomes a bag of visual words.
02 · WHAT IT SAYS
Read every frame.
OCR pulls the on-screen text, and Lucene scores how much two videos’ words overlap.
03 · TOGETHER
Combine. Then rank.
Visual and textual similarity are blended 0.8 / 0.2, and text joins in only when it can actually tell bugs apart.
EXISTING REPORTS · RANKED
Evaluated on six Android apps
0
duplicate-detection tasks
0%
of tasks rank a true duplicate in the top 2
67.7%ranked a duplicate first
79.8%mean reciprocal rank
65.1%of duplicate-hunting time saved
It Takes Two to Tango: Combining Visual and Textual Information for Detecting Duplicate Video-Based Bug Reports
Nathan Cooper · Carlos Bernal-Cárdenas · Oscar Chaparro · Kevin Moran · Denys Poshyvanyk
ICSE 2021 · IEEE/ACM International Conference on Software Engineering