SE(3) rigid transform class that allows compounding of 6-DOF poses and provides common transformations that are commonly seen in geometric problem

Vision-based 3D box container. The tvec is computed from projected center, depth, and intrinsics.

Vanilla FPN and PAN

Implement test-time augmentation for detection data. It is a callable which takes a dataset dict from a detection dataset, and returns a

Adapted from detectron2: https://github.com/facebookresearch/detectron2/blob/master/detectron2/structures/image_list.py) Key differe

A class for bird's eye view visualization, which generates a canvas of bird's eye view image, The class concerns two types of transformations:

Adapted from detectron2: detectron2.utils.visualizer Key difference: load inference results on disk generated by COCOEvaluator

This is adapted from: https://github.com/facebookresearch/detectron2/blob/master/detectron2/data/dataset_mapper.py The changes from

Patch DetectionEval of NuScenes devkit to only evaluate on samples in the predictions.

A MBV2 block with fused depthwise and channelwise convolutions.

Intersetion Over Union (IoU) loss which supports three different IoU computations: * IoU * Linear IoU * gIoU

Assumptions: 1) The dataset consists of in-order groups, i.e. [*group-1-items, *group-2-items, ...] 2) In the dataloader, per-gpu

This module is used in RetinaNet to generate extra layers, P6 and P7 from C5 feature.

This module is used in FCOS to generate extra layers

A MobileNetV2 block as used in the MobileNetV2 architecture.

In addition to 2D / 3D boxes, each instance also has attribute and speed. Assumption: image transformation does not change attributes and sp

If the checkpoint (`config.model.CKPT`) is an S3 path, then downloaded it and replace the path with local path.

A GeneralizedRCNN with test-time augmentation enabled. Its :meth:`__call__` method has the same interface as :meth:`GeneralizedRCNN.forward`.

Dialable batch-norm layers. Typical use case: all FPN layers shares a 2D convolutional decoder, but the batch-norm layers are not shared. Tha

Dialable modules. Typically used with hierarchical output from FPN feature extractors. Separate modules are applied to each FPN layer.

Separable Conv

If W&B is enabled, then 1) initialize W&B, 2) derive the path of output directory using W&B ID, and 3) set it as hydra working