Upload README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-nc-sa-4.0
|
| 3 |
+
tags:
|
| 4 |
+
- point-cloud
|
| 5 |
+
- semantic-segmentation
|
| 6 |
+
- instance-segmentation
|
| 7 |
+
- pointcept
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# Checkpoints for the submission "relabelling voxel indices"
|
| 11 |
+
|
| 12 |
+
Weights of the models we trained for the paper, to be read with the code release (an overlay
|
| 13 |
+
on [Pointcept](https://github.com/Pointcept/Pointcept) `v1.7.0`). Every model here was trained
|
| 14 |
+
**from scratch** with Pointcept's unmodified model code; no third-party weights are included.
|
| 15 |
+
|
| 16 |
+
Layout: `exp/<dataset>/<config>/model/model_best.pth` plus the run's `train.log`. The directory
|
| 17 |
+
name is the Pointcept configuration that trained it; a second seed ends in `-seed1`. Each file
|
| 18 |
+
holds the model `state_dict` (plus `epoch` and `best_metric_value`); optimizer state is removed.
|
| 19 |
+
`SHA256SUMS` lists the hashes. The published LitePT-S checkpoints the paper also reads are not
|
| 20 |
+
copied here: they are on [prs-eth/LitePT](https://huggingface.co/prs-eth/LitePT).
|
| 21 |
+
|
| 22 |
+
| dataset | checkpoints |
|
| 23 |
+
|---|---|
|
| 24 |
+
| ScanNet v2 | `semseg-litept-v1m1-0-small`, `semseg-litept-v1m1-0b-small-scaleaug`, `semseg-litept-v1m1-0-small-train3tau`, `semseg-litept-v1m2-0-small` (+ `-seed1`), `semseg-pt-v3m1-0-base`, `semseg-spunet-v1m1-0-base`, `semseg-pt-v2m2-0-base`, `insseg-pointgroup-v1m1-0-spunet-base` |
|
| 25 |
+
| nuScenes | `semseg-litept-v1m1-0-small` (+ `-seed1`), `semseg-litept-v1m1-0b-small-scaleaug` (+ `-seed1`), `semseg-litept-v1m2-0-small`, `semseg-litept-v1m2-0b-small-scaleaug`, `semseg-pt-v3m1-0-base`, `semseg-spunet-v1m1-0-base` |
|
| 26 |
+
|
| 27 |
+
## Licence and data terms
|
| 28 |
+
|
| 29 |
+
These weights are released for **non-commercial research only**, under
|
| 30 |
+
[CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/), because of the data
|
| 31 |
+
they were trained on:
|
| 32 |
+
|
| 33 |
+
* **nuScenes** is provided by Motional under CC BY-NC-SA 4.0 and the
|
| 34 |
+
[nuScenes Terms of Use](https://www.nuscenes.org/terms-of-use). The nuScenes checkpoints are
|
| 35 |
+
shared under the same licence. This release is not endorsed by Motional.
|
| 36 |
+
Caesar et al., *nuScenes: A multimodal dataset for autonomous driving*, CVPR 2020.
|
| 37 |
+
* **ScanNet** may be used for non-commercial research and educational purposes only, under
|
| 38 |
+
the [ScanNet Terms of Use](https://kaldir.vc.in.tum.de/scannet/ScanNet_TOS.pdf). No ScanNet
|
| 39 |
+
data is included here. Dai et al., *ScanNet: Richly-annotated 3D Reconstructions of Indoor
|
| 40 |
+
Scenes*, CVPR 2017.
|
| 41 |
+
|
| 42 |
+
The training and model code (Pointcept, LitePT) is MIT-licensed and is not included here.
|