search

Qiuzhongyuan / pointpillars_pytorch_trt

Apache License 2.0
77 stars 24 forks source link

pointpillars_large_model.onnx model shape means（ inputs & outputs ） #7

Closed Ryze0124 closed 3 years ago

Ryze0124 commented 3 years ago

pointpillars_large_model.onnx model has two outputs (pointpillars_output1 & pointpillars_output2), can you help me to explain their shape mesns（outputs all is 1 x 100 x 9）， and VoxelGeneratorV1 is dynamic plugin，but onnx model input is fixed （25000 x 5），why？

Qiuzhongyuan commented 3 years ago
  1. 'nms_post_maxsize' in NMS layer is 100, so the outputs' dim 1 is 100, 9 means xyzlwhr+score + cls_idx for per box;
  2. pcs are different between frames, so we must set a big shape for all the frames. in kitti reduced pcs, 25000 is enough,the 'valid' tensor will be used to filter the data.
  3. all the fixed shape is used to deploy model with tensorrt more easily