Hi, thanks for releasing EventBench.
I noticed several event_ocr samples whose ground-truth target is an empty string, but the stacked event visualization appears to contain readable text. This can incorrectly reward blank OCR predictions.
Examples:
| File |
Current target |
Visible text |
event_ocr/recording_2023-03-09_19-05-52.npz |
"" |
SPA |
event_ocr/recording_2023-03-09_14-11-35.npz |
"" |
理：摩托车 电动车 档位车 |
event_ocr/recording_2023-04-01_15-15-17.npz |
"" |
浦电路155号 |
event_ocr/recording_2023-03-09_14-19-39.npz |
"" |
巨鳌轴承 |
Could you please check whether these OCR annotations should be updated?
Hi, thanks for releasing EventBench.
I noticed several
event_ocrsamples whose ground-truth target is an empty string, but the stacked event visualization appears to contain readable text. This can incorrectly reward blank OCR predictions.Examples:
event_ocr/recording_2023-03-09_19-05-52.npz""SPAevent_ocr/recording_2023-03-09_14-11-35.npz""理：摩托车 电动车 档位车event_ocr/recording_2023-04-01_15-15-17.npz""浦电路155号event_ocr/recording_2023-03-09_14-19-39.npz""巨鳌轴承Could you please check whether these OCR annotations should be updated?