Dataset statistics and missing images

[tjubiit]

Dataset statistics and missing images

[Hanqing-Sun]

Thank you for your interest in our dataset. We did some investigations and apologize for the late reply.

1. The duplication is confirmed using md5sum. 500+ dups are found and one of them is cross-split. We suppose that won't impact the training and benchmarking too much, please let us know if you observe significant performance issues.
2. The number of images reported in the paper is 115,354 (traffic 60,266 + campus 55,088, Tab. I). Discarding all the dups won't decrease it to the commented 105,076. Do we share an identical stat dimension or did we miss anything?
3. We find the files in the dhd-traffic annotation instead of that in the dhd-campus, therefore, the path should be dhd_traffic/images/train/1499435044634.jpg. The md5s of all files are provided in the README.md, you can have a check of the downloaded zips.
4. The pedstrian is a typo, we will fix it later.