I learned that the radar sensor you are using for the recordings can give so called cluster information and also ‘regular’ object lists. In the data set I can only find the cluster information. Is it also possible to get the object information (i.e. including tracking)?
Hi, we provide these outputs: https://github.com/nutonomy/nuscenes-devkit/blob/master/python-sdk/nuscenes/utils/data_classes.py#L313
Unfortunately, these are only cluster states, dynamic properties and false alarm probabilities. There is no real “object” information.
Thanks for the quick reply. I have a question that is partly a follow up question:
When looking at the unfiltered radar points (including such with high false positive probability and different invalid states etc.), I noticed that the total of the unfiltered points is always 125. That makes sense as probably the maximum of clusters the sensor can provide is 125. I am, however, wondering if those 125 cluster points are always the ones detected at that point in time. What happens if the sensor does not have 125 detections in one frame? Does it keep old stages and how can they be told appard? If it has more than 125 detections, I guess there is some internal logic to cut them off?
I would assume the radar returns at time t are independent of time t-1, but I am not sure. You could try to render a video at 12Hz and see if old detections are kept in memory.
Perhaps the manuals I sent you can tell you more?
I am working with the unfiltered radar points too, so I disabled all filters.
Please have a look at the end of scene-0757. At the, for example, for the 39th sample I just have 118 points. Using the 30th sample I receive 112 points.
thanks for letting me know. I guess my sample size of scenes I tried was too small. I ran it with scene-0757 and also with scene-0758 for testing and also noticed that in some cases the number of unfiltered radar points is smaller than 125. For your particular example, however, I got 125 in the 39th sample and also for the 30th sample. But I got, e.g., 118 for the 40th sample and 119 for the 41st.
I guess the sensor will in most cases detect some clutter in the directions it is shooting rays which than leads to 125 “detections” in most cases. Probably many cases then have a low RCS and are rated invalid afterwards.