The evalutation code missing the matchig part to determine the right order of multi-person results. Maybe carry out grid search to find the best matched person with lowest error?
Have a good day!
Maybe a simpler option is to use the first frame bounding box to match the id during the inference. In the first frame in 3dpw, by design, all the persons to recognize are clearly visible.
This doesn't require changing the evaluation code which is somewhat safer in such a late stage
The website has been updated to clear all ambiguities. The first frame GT data can be used to ascertain the number and identity of people tracked
Posted by: aymen @ July 17, 2020, 10:06 p.m.