OpenMMLab v2.0 Refactoring

In this version, we make large refactoring based on MMEngine to achieve unified data elements, model interfaces, visualizers, evaluators and other runtime modules across different datasets, tasks and even codebases. A brief summary for this refactoring is as follows:

  • Data Element:

    • We add Det3DDataSample as the common data element passing through datasets and models. It inherits from DetDataSample in mmdetection and implement InstanceData, PixelData, and LabelData inheriting from BaseDataElement in MMEngine to represent different types of ground truth labels or predictions.

  • Datasets:

    • We add Det3DDataset and Seg3DDataset as the base datasets to inherit from the unified BaseDataset in MMEngine. They implement most functions that are commonly used across different datasets and simplify the info loading/processing in the current datasets. Re-defined input arguments and functions can be most re-used in different datasets, which are important for the implementation of customized datasets.

    • We define the common keys across different datasets and unify all the info files with a standard protocol. The same info is more clear for users because they share the same key across different dataset infos. Besides, for different settings, such as camera-only and LiDAR-only methods, we no longer need different info formats (like the previous pkl and json files). We can just revise the parse_data_info to read the necessary information from a complete info file.

    • We add train_dataloader, val_dataloader and test_dataloader to replace the original data in the config. It simplify the levels of data-related fields.

  • Data Transforms

    • Based on the basic transforms and wrappers re-implemented and simplified in the latest MMCV, we refactor data transforms to inherit from them.

    • We also adjust the implementation of current data pipelines to make them compatible with our latest data protocol.

    • Normalization, padding of images and voxelization operations are moved to the data-preprocessing.

    • DefaultFormatBundle3D and Collect3D are replaced with PackDet3DInputs to pack the data into the element format as model input.

  • Models

    • Unify the model interface as inputs, data_samples, return_loss=False

    • The basic pre-processing before model forward includes: 1) convert input from CPU to GPU tensors; 2) padding images; 3) normalize images; 4) voxelization.

    • Return loss_dict during training while return list[data_sample] during inference

    • Simply function interfaces in the models

    • Add preprocess_cfg in the model configs for pre-processing

  • Visualizer

  • Evaluator

    • Decouple evaluators from datasets to make them more flexible: the evaluation codes of each dataset are implemented as a metric class exclusively.

    • Add evaluator information to the current dataset configs

  • Registry

    • Refactor all the registries to inherit from root registries in MMEngine

    • When using modules from other codebases, it is necessary to specify the registry scope, such as mmdet.ResNet

  • Others: Refactor logging, hooks, scheduler, runner and other runtime configs based on MMEngine


Operators Migration

We have adopted CUDA operators compiled from mmcv and removed all the CUDA operators in mmdet3d. We now do not need to compile the CUDA operators in mmdet3d anymore.

Waymo dataset converter refactoring

In this version we did a major code refactoring that boosted the performance of waymo dataset conversion by multiprocessing. Meanwhile, we also fixed the imprecise timestamps saving issue in waymo dataset conversion. This change introduces following backward compatibility breaks:

  • The point cloud .bin files of waymo dataset need to be regenerated. In the .bin files each point occupies 6 float32 and the meaning of the last float32 now changed from imprecise timestamps to range frame offset. The range frame offset for each point is calculated asri * h * w + row * w + col if the point is from the TOP lidar or -1 otherwise. The h, w denote the height and width of the TOP lidar’s range frame. The ri, row, col denote the return index, the row and the column of the range frame where each point locates. Following tables show the difference across the change:


Element offset (float32) 0 1 2 3 4 5
Bytes offset 0 4 8 12 16 20
Meaning x y z intensity elongation imprecise timestamp


Element offset (float32) 0 1 2 3 4 5
Bytes offset 0 4 8 12 16 20
Meaning x y z intensity elongation range frame offset
  • The objects’ point cloud .bin files in the GT-database of waymo dataset need to be regenerated because we also dumped the range frame offset for each point into it. Following tables show the difference across the change:


Element offset (float32) 0 1 2 3 4
Bytes offset 0 4 8 12 16
Meaning x y z intensity elongation


Element offset (float32) 0 1 2 3 4 5
Bytes offset 0 4 8 12 16 20
Meaning x y z intensity elongation range frame offset
  • Any configuration that uses waymo dataset with GT Augmentation should change the db_sampler.points_loader.load_dim from 5 to 6.


Coordinate system refactoring

In this version, we did a major code refactoring which improved the consistency among the three coordinate systems (and corresponding box representation), LiDAR, Camera, and Depth. A brief summary for this refactoring is as follows:

  • The three coordinate systems are all right-handed now (which means the yaw angle increases in the counterclockwise direction).

  • The LiDAR system (x_size, y_size, z_size) corresponds to (l, w, h) instead of (w, l, h). This is more natural since l is parallel with the direction where the yaw angle is zero, and we prefer using the positive direction of the x axis as that direction, which is exactly how we define yaw angle in Depth and Camera coordinate systems.

  • The APIs for box-related operations are improved and now are more user-friendly.


Since definitions of box representation have changed, the annotation data of most datasets require updating:

  • SUN RGB-D: Yaw angles in the annotation should be reversed.

  • KITTI: For LiDAR boxes in GT databases, (x_size, y_size, z_size, yaw) out of (x, y, z, x_size, y_size, z_size) should be converted from the old LiDAR coordinate system to the new one. The training/validation data annotations should be left unchanged since they are under the Camera coordinate system, which is unmodified after the refactoring.

  • Waymo: Same as KITTI.

  • nuScenes: For LiDAR boxes in training/validation data and GT databases, (x_size, y_size, z_size, yaw) out of (x, y, z, x_size, y_size, z_size) should be converted.

  • Lyft: Same as nuScenes.

Please regenerate the data annotation/GT database files or use to update the data.

To use boxes under Depth and LiDAR coordinate systems, or to convert boxes between different coordinate systems, users should be aware of the difference between the old and new definitions. For example, the rotation, flipping, and bev functions of DepthInstance3DBoxes and LiDARInstance3DBoxes and box conversion functions have all been reimplemented in the refactoring.

Consequently, functions like output_to_lyft_box undergo small modification to adapt to the new LiDAR/Depth box.

Since the LiDAR system (x_size, y_size, z_size) now corresponds to (l, w, h) instead of (w, l, h), the anchor sizes for LiDAR boxes are also changed, e.g., from [1.6, 3.9, 1.56] to [3.9, 1.6, 1.56].

Functions only involving points are generally unaffected except if they rely on some refactored utility functions such as rotation_3d_in_axis.

Other BC-breaking or new features:

  • array_converter: Please refer to Functions wrapped with array_converter can convert array-like input types of torch.Tensor, np.ndarray, and list/tuple/float to torch.Tensor to process in an unified PyTorch pipeline. The result may finally be converted back to the input type. Most functions in are wrapped with array_converter.

  • points_in_boxes and points_in_boxes_batch will be deprecated soon. They are renamed to points_in_boxes_part and points_in_boxes_all respectively, with more detailed docstrings. The major difference of the two functions is that if a point is enclosed by multiple boxes, points_in_boxes_part will only return the index of the first enclosing box while points_in_boxes_all will return all the indices of enclosing boxes.

  • rotation_3d_in_axis: Please refer to Now this function supports multiple input types and more options. The function with the same name in is deleted since we do not need another function to tackle with NumPy data. rotation_2d, points_cam2img, and limit_period in are also deleted for the same reason.

  • bev method of CameraInstance3DBoxes: Changed it to be consistent with the definition of bev in Depth and LiDAR coordinate systems.

  • Data augmentation utils in now follow the rules of a right-handed system.

  • We do not need the yaw hacking in KITTI anymore after refining get_direction_target. Interested users may refer to PR #677 .


Returned values of QueryAndGroup operation

We modified the returned grouped_xyz value of operation QueryAndGroup to support PAConv segmentor. Originally, the grouped_xyz is centered by subtracting the grouping centers, which represents the relative positions of grouped points. Now, we didn’t perform such subtraction and the returned grouped_xyz stands for the absolute coordinates of these points.

Note that, the other returned variables of QueryAndGroup such as new_features, unique_cnt and grouped_idx are not affected.

NuScenes coco-style data pre-processing

We remove the rotation and dimension hack in the monocular 3D detection on nuScenes. Specifically, we transform the rotation and dimension of boxes defined by nuScenes devkit to the coordinate system of our CameraInstance3DBoxes in the pre-processing and transform them back in the post-processing. In this way, we can remove the corresponding hack used in the visualization tools. The modification also guarantees the correctness of all the operations based on our CameraInstance3DBoxes (such as NMS and flip augmentation) when training monocular 3D detectors.

The modification only influences nuScenes coco-style json files. Please re-run the nuScenes data preparation script if necessary. See more details in the PR #744.

ScanNet dataset for ImVoxelNet

We adopt a new pre-processing procedure for the ScanNet dataset in order to support ImVoxelNet, which is a multi-view method requiring image data. In previous versions of MMDetection3D, ScanNet dataset was only used for point cloud based 3D detection and segmentation methods. We plan adding ImVoxelNet to our model zoo, thus updating ScanNet correspondingly by adding image-related pre-processing steps. Specifically, we made these changes:

  • Add script for extracting RGB data.

  • Update script for annotation creating.

  • Add instructions in the documents on preparing image data.

Please refer to the ScanNet for more details.


MMCV Version

In order to fix the problem that the priority of EvalHook is too low, all hook priorities have been re-adjusted in 1.3.8, so MMDetection 2.14.0 needs to rely on the latest MMCV 1.3.8 version. For related information, please refer to #1120, for related issues, please refer to #5343.

Unified parameter initialization

To unify the parameter initialization in OpenMMLab projects, MMCV supports BaseModule that accepts init_cfg to allow the modules’ parameters initialized in a flexible and unified manner. Now the users need to explicitly call model.init_weights() in the training script to initialize the model (as in here, previously this was handled by the detector. Please refer to PR #622 for details.


We modified the dataset augmentation function BackgroundPointsFilter(here). In previous version of MMdetection3D, BackgroundPointsFilter changes the gt_bboxes_3d’s bottom center to the gravity center. In MMDetection3D 0.15.0, BackgroundPointsFilter will not change it. Please refer to PR #609 for details.

Enhance IndoorPatchPointSample transform

We enhance the pipeline function IndoorPatchPointSample used in point cloud segmentation task by adding more choices for patch selection. Also, we plan to remove the unused parameter sample_rate in the future. Please modify the code as well as the config files accordingly if you use this transform.


Dataset class for 3D segmentation task

We remove a useless parameter label_weight from segmentation datasets including Custom3DSegDataset, ScanNetSegDataset and S3DISSegDataset since this weight is utilized in the loss function of model class. Please modify the code as well as the config files accordingly if you use or inherit from these codes.

ScanNet data pre-processing

We adopt new pre-processing and conversion steps of ScanNet dataset. In previous versions of MMDetection3D, ScanNet dataset was only used for 3D detection task, where we trained on the training set and tested on the validation set. In MMDetection3D 0.14.0, we further support 3D segmentation task on ScanNet, which includes online benchmarking on test set. Since the alignment matrix is not provided for test set data, we abandon the alignment of points in data generation steps to support both tasks. Besides, as 3D segmentation requires per-point prediction, we also remove the down-sampling step in data generation.

  • In the new ScanNet processing scripts, we save the unaligned points for all the training, validation and test set. For train and val set with annotations, we also store the axis_align_matrix in data infos. For ground-truth bounding boxes, we store boxes in both aligned and unaligned coordinates with key gt_boxes_upright_depth and key unaligned_gt_boxes_upright_depth respectively in data infos.

  • In ScanNetDataset, we now load the axis_align_matrix as a part of data annotations. If it is not contained in old data infos, we will use identity matrix for compatibility. We also add a transform function GlobalAlignment in ScanNet detection data pipeline to align the points.

  • Since the aligned boxes share the same key as in old data infos, we do not need to modify the code related to it. But do remember that they are not in the same coordinate system as the saved points.

  • There is an PointSample pipeline in the data pipelines for ScanNet detection task which down-samples points. So removing down-sampling in data generation will not affect the code.

We have trained a VoteNet model on the newly processed ScanNet dataset and get similar benchmark results. In order to prepare ScanNet data for both detection and segmentation tasks, please re-run the new pre-processing scripts following the ScanNet


SUNRGBD dataset for ImVoteNet

We adopt a new pre-processing procedure for the SUNRGBD dataset in order to support ImVoteNet, which is a multi-modality method requiring both image and point cloud data. In previous versions of MMDetection3D, SUNRGBD dataset was only used for point cloud based 3D detection methods. In MMDetection3D 0.12.0, we add ImVoteNet to our model zoo, thus updating SUNRGBD correspondingly by adding image-related pre-processing steps. Specifically, we made these changes:

  • Fix a bug in the image file path in meta data.

  • Convert calibration matrices from double to float to avoid type mismatch in further operations.

  • Add instructions in the documents on preparing image data.

Please refer to the SUNRGBD for more details.


VoteNet and H3DNet model structure update

In MMDetection 0.6.0, we updated the model structures of VoteNet and H3DNet, therefore model checkpoints generated by MMDetection < 0.6.0 should be first converted to a format compatible with the latest structures via and . For more details, please refer to the VoteNet and H3DNet

Read the Docs v: latest
On Read the Docs
Project Home

Free document hosting provided by Read the Docs.