Pointcept Versions Save

Changelog:

1. Release most configs for PTv3 indoor perception. (I forget the ScanNet200 PPT version)
2. Release several configs for PTv3 outdoor perception. (Computing resources have been limited recently, I want to make sure the final version of the code is reproducible)
3. Fix the processed NuScenes info download link.
4. Refine Octformer import logits
5. Refine PPT

Tasks:

1. All configs for PTv3
2. Support ImageNet classification
3. Support object-level dataset
4. Pre-training technology on object-level datasets

(Hope I will have enough time to finish these tasks)

Changelog:

1. Release model code for Point Transformer V3;
2. Release configs and exp records for training scratched PTv3 on ScanNet and ScanNet200;
3. Fix displacement computation for Swin3D and OctFormer;
4. Increase timeout limitation to 60min;
5. More flexible import strategy, make most model default Imported;
6. Add Point. PointModule. PointSequence;
7. Detached PDNorm as a Pointcept Module;
8. Add DefaultSegmentorV2 (Backbone only extract feature)

1. Release PPT (check MultiDatasetLoader, Trainer and model);
2. Add support to Waymo;
3. Optimize precise test (dual grid sampling, refer NuScenes config);
4. Rename discrete_coord to grid_coord;
5. Add test script for S3DIS 6-fold cross-validation;

1. Code release of CAC;
2. Adding support to NuScenes and SemanticKITTI with PTv2;
3. Adding support to Swin3d and OctFormer
4. Adding support to Structured3D pretraining

1. Release code for Masked Scene Contrast (MSC);
2. Add support to PointContrast and Contrastive Scene Contexts;
3. Move pointcept.utils.losses to pointcept.models.losses
4. Rename Voxelize to GridSample;
5. Remove the max_batch_point mechanism;
6. Enable PreciseEvaluator as a default hook for semantic segmentation;
7. Enable DDP testing (hook and test script).

1. Adding support to instance segmentation with PointGroup.
2. Adding support to Python >= 3.10.
3. Unifing config and model naming.
4. Adding PreciseEvaluator as a hook, automatically operating the testing process after train.

We upgraded the released codebase of Point Transformer V2 and named it Pointcept. Compared with the PTv2 released code, the main change is as follows:

1. Enable the hook mechanism, which widely increases the flexibility of Pointcept.
2. Move loss computation from trainer to model, and after that, Pointcept can support various downstream tasks.
3. Better dataflow design, to distinguish labels for different tasks, we use category for classification label, segment for semantic segmentation, instance for instance segmentation, bbox for object detection.
4. Adding support to ModelNet40.
5. Adding support to ScanNet Data Efficient Benchmark.
6. Enable PyTorch Profiler and Share Memory as a hook.
7. Better detail design.