This repository provides scripts to convert the weights of MegaDetector 5.0 into the ONNX and TensorRT formats to simplify the usage in embedded devices. yolort further optimizes the runtime performance by embedding the pre- and post-processing steps into the models.
Already converted weights can be found under releases.
- enable batch sizes larger than 1 and flexible spatial dimensions for the TensorRT models
- perform benchmarks on some embedded devices
- automatically compare detection performance of derived models
The test.jpg file is provided by the Caltech Camera Traps dataset under the terms of the Community Data License Agreement (permissive variant).