---
license: other
tags:
  - sdtagnet
  - nuscenes
  - hd-map
  - online-mapping
  - autonomous-driving
  - new-split
  - long-range
library_name: mmdetection3d
---

# SDTagNet - nuScenes (new split, long range), ResNet-50, 30ep schedule

Best checkpoints from an SDTagNet training run on nuScenes with the **new split** and the **100 x 50 m long-range setting** (`pc_range = [-25, -50, -10, 25, 50, 10]`, BEV `(200, 100)`). ResNet-50 image backbone, camera-only SDTagNet, 30-epoch schedule.

This repository replaces the earlier upload under the same repo name with the later `sdtagnet_nusc_r50_30ep_100_50_range_mapping_last` run.

## Evaluation on long-range new-split val

| Checkpoint | Save time (UTC) | NuscMap_chamfer/mAP | divider_AP | boundary_AP | ped_crossing_AP |
| --- | --- | --- | --- | --- | --- |
| `mAP_epoch_8.pth` | 2026-04-29 22:15 | 0.2070 | 0.2399 | 0.2700 | 0.1112 |
| `mAP_epoch_20.pth` | 2026-04-30 19:49 | **0.2251** | 0.2490 | 0.3077 | 0.1188 |

AP by chamfer threshold (`@1.0 / @1.5 / @2.0 m`):

### `mAP_epoch_8.pth`

| Class | @1.0 | @1.5 | @2.0 |
| --- | --- | --- | --- |
| divider | 0.1364 | 0.2490 | 0.3343 |
| boundary | 0.1064 | 0.2837 | 0.4197 |
| ped_crossing | 0.0349 | 0.1112 | 0.1876 |

### `mAP_epoch_20.pth`

| Class | @1.0 | @1.5 | @2.0 |
| --- | --- | --- | --- |
| divider | 0.1534 | 0.2587 | 0.3348 |
| boundary | 0.1293 | 0.3199 | 0.4737 |
| ped_crossing | 0.0368 | 0.1187 | 0.2009 |

Recorded mAP progression:

| Epoch | mAP |
| --- | --- |
| 2 | 0.1345 |
| 4 | 0.1855 |
| 6 | 0.1965 |
| 8 | 0.2070 |
| 10 | 0.2157 |
| 12 | 0.2180 |
| 14 | 0.2240 |
| 16 | 0.2200 |
| 18 | 0.2238 |
| **20** | **0.2251** |
| 22 | 0.2210 |

## Range / BEV / modality

- `pc_range = [-25.0, -50.0, -10.0, 25.0, 50.0, 10.0]`
- `bev_size = (200, 100)`
- `use_lidar = False` (camera-only SDTagNet)
- Image backbone: ResNet-50
- Training schedule in the saved config: 30 epochs

## Files

| Path | Description |
| --- | --- |
| `mAP_epoch_8.pth` | Best checkpoint at epoch 8 (~579 MB) |
| `mAP_epoch_20.pth` | Best checkpoint at epoch 20, highest recorded mAP in this run (~579 MB) |
| `sdtagnet_nusc_r50_30ep_100_50_range_mapping_last.py` | Exact config snapshot saved in the work_dir |
| `logs/20260429_103003.log` / `.log.json` | Training/eval log covering epochs 2-8 |
| `logs/20260430_021112.log` / `.log.json` | Continuation log covering epochs 10-22 |

## Config notes

The config expects:

- `data_root = '/datasets/public/nuscenes/'`
- train/val pkl: `nuscenes_map_infos_temporal_newsplit_100x50_{train,val}.pkl`
- eval json: `nuscenes_map_anns_val_sdtagnet_100x50.json`
- `work_dir = './work_dirs/sdtagnet_nusc_r50_30ep_100_50_range_mapping_last'`

Override paths for your own environment.

## Related checkpoints

- `sung98/sdtagnet-nusc-r50-110ep-newsplit` - SDTagNet, standard range, new split
- `sung98/sdtagnet-nusc-r50-30ep-oldsplit` - SDTagNet, old split, standard range
- `sung98/maptrv2-nusc-r50-110ep-lidar-newsplit` - MapTRv2 LiDAR, new split
- `sung98/maptrv2-nusc-r50-24ep-lidar-oldsplit` - MapTRv2 LiDAR, old split
- `sung98/maptrv2-nusc-r50-24ep-lidar-newsplit-longrange` - MapTRv2 LiDAR, long range, new split

Numbers across runs are not directly comparable (different range / split / eval thresholds / model).

## Loading

```python
from mmcv import Config
from mmdet3d.models import build_model
from mmcv.runner import load_checkpoint

cfg = Config.fromfile('sdtagnet_nusc_r50_30ep_100_50_range_mapping_last.py')
cfg.model.train_cfg = None
model = build_model(cfg.model, test_cfg=cfg.get('test_cfg'))
load_checkpoint(model, 'mAP_epoch_20.pth', map_location='cpu')
model.eval()
```