Is there bug in NMSPlugin?

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

As far as I understood NMSPlugin has two outputs:

  • object parameters table - calss, confidence, box and so on
  • number of actual items in object parameters table

Second output has the only integer value, and it`s supposed to be so always.

So when I set network precision in FT32 - everything is fine, but with same input data in FT16 mode - second output contains garbage. nevertheless first input contains correct parameters table.

Checked on TensorRT 5.0 GA and 5.1 RC

2

hello, to help us debug, can you please share a small repro with model and detectionOutput/nms plugin code (fp32, fp16) that exhibit the symptoms you are seeing?