running https://github.com/NVIDIA-AI-IOT/tf_trt_models/blob/master/examples/detection/detection.ipynb in nvidia/tensorflow:19.01-p3 container with Tesla V100 16GB:
Running the notebook with
MODEL = 'ssd_mobilenet_v2_coco'
and evaluating with “FP32” and “FP16”:
in both cases, 'Average runtime" will be at aroung 14-15ms, not showing any significant improvements.
Running provided samples for TensorRT optimization for image classification works just fine, improving inderence time by ~3-4 times.
Does TensorRT have some issues with object detection that are not apparent in plain image classification?