---
license: apache-2.0
tags:
  - autonomous-driving
  - hd-map
  - bev-perception
  - nuscenes
  - multi-modal-fusion
  - satellite
---

# ResilientMap — Stage 1 (Pixel-Gate Fusion), nuScenes old-split

Checkpoints and training artifacts for **Stage 1** of the ResilientMap project:
camera + satellite BEV fusion via a confidence-weighted pixel gate, trained on
the nuScenes **old split** (full data).

This is the BEV-feature pre-training stage. The vectorized HD map decoder
(MMQuery, Stage 2/3) is trained on top of these BEVs and is released
separately.

## Files

| File | Size | Notes |
|---|---|---|
| `iter_55932.pth` | 897 MB | **Final checkpoint** (latest, recommended) |
| `iter_37288.pth` | 897 MB | Intermediate checkpoint (4-th of 6 saves) |
| `satmaptracker_pixelgate_stage1_full.py` | 23 KB | Full training/eval config |
| `20260417_174220.log` | 862 KB | Training log (text) |
| `20260417_174220.log.json` | 857 KB | Training log (JSON, per-iter metrics) |
| `tf_logs/` | — | TensorBoard event files |

> Earlier intermediate checkpoints (`iter_9322`, `iter_18644`, `iter_27966`,
> `iter_46610`) were not uploaded to keep the repo lean. The `latest.pth`
> symlink in the original work-dir pointed to `iter_55932.pth`.

## Architecture (one paragraph)

A BEVFormer (ResNet-50 + FPN + TemporalSelfAttention) camera branch produces a
`(256, 50, 100)` BEV. A separate ResNet-50 + FPN satellite encoder produces an
ego-aligned satellite BEV at the same shape. Two frozen segmentation decoders
output per-pixel confidences for each modality, and the BEVs are combined as

```
g(x,y)    = cam_conf(x,y) / (cam_conf(x,y) + sat_conf(x,y) + eps)
fused_bev = g * cam_bev + (1 - g) * sat_bev
```

This residual-free, confidence-weighted gate gives the fused stream a clean
fall-back to whichever modality is more reliable per pixel.

## Headline numbers (nuScenes new-split, 1/3 data, BEV mIoU; this checkpoint is the **old-split / full-data** counterpart, retraining of the same recipe)

| Condition | Fused | Cam-only | Sat-only |
|---|---|---|---|
| Clean | **0.512** | 0.452 | 0.403 |
| 6 cameras zeroed | **0.381** | 0.013 | 0.403 |
| Satellite zeroed | **0.436** | 0.452 | 0.000 |

Per-camera importance (impact when a single camera is zeroed):
`FRONT -0.058 >> BACK -0.027 > FRONT_LEFT/RIGHT -0.018/-0.015 > BACK_LEFT/RIGHT -0.006`.

## Intended use

Research checkpoint for online HD map construction with multi-modal sensor
redundancy. Not production-ready.

## Citation

Citation entry will be added when the accompanying paper is released.