Mono3D Release Fix

As pointed out in : #4

The originally provided Onnx file for object 3d detection does not work as expected. Here is the fix for it.

Metric 3D Release

Export the metric3d_vit_small version of Metric3D into ONNX model.

Important changes and improvements over the base model:

• Use base attention instead of xFormers during compilation.
• Use bilinear instead of bicubic interpolation for patch embedding. So that the model can run on TensorRT, tested on Jetpack 4.6 Jetson Xavier / x86 (even though the speed is not perfect on Xavier, the model runs fast on modern 60-series GPUs).
• Projecting depth images into 3D and also allowing performing coordinate transform within the ONNX model. Minimize the operation needed on the CPU for metric3D models, tested on Jetson TensorRT settings.
• The exportation scripts is in tool_boxes/export_metric_3d.ipynb

What's Changed

• Metric3d by @Owen-Liuyuxuan in #3

New Contributors

• @Owen-Liuyuxuan made their first contribution in #3

Full Changelog: v1.0.1...v1.1

v1.0.1 DLA34 Deformable Convolutional MonoFlex model

Manage to export Deformable convolutions with torchvision and exporter .

The DLA3D deformable convolutional module performs significantly better than the YOLOX backbone.

v1.0 YoloXMonoFlex, BisenetV1, Res101-FSNet

YoloXMonoFlex: MonoFlex trained with multiple datasets from vision_factory with YOLOX backbone.

BisenetV1: Segmentation trained with multiple datasets from vision_factory with BisenetV1 backbone.

Res101-FSNet: FSNet exported to ONNX format from the original release

All these are done in the development branch