# S23DR 2026 — WireframeDETR

Submission to the [S23DR 2026 Challenge](https://huggingface.co/spaces/usm3d/S23DR2026).

**Public test HSS: 0.575** (F1=0.664, IoU=0.516)

> This project was built on Modal GPU credits left over from a previous hackathon. Each run took hours, so the budget kept experiments deliberate. Several training configurations were explored, but a full factorial study of every contribution wasn't feasible.


## Approach

End-to-end 3D wireframe prediction via DETR-style set prediction over COLMAP point clouds.
Each predicted edge is a 6D coordinate pair `(x1,y1,z1,x2,y2,z2)` regressed by a learned query.
Hungarian matching assigns predictions to ground-truth edges at training time.

**Our contributions:**
- **Contrastive Denoising Training (CDN)** — adapted from [DN-DETR](https://arxiv.org/abs/2203.01305); injects GT-aligned denoising queries alongside learned queries to stabilise Hungarian matching in early epochs
- **Multi-scale encoder** — learned softmax-weighted average of last K=3 encoder layer outputs, giving the decoder access to both fine-grained and abstract representations
- **Progressive auxiliary loss weighting** — decoder layer i weighted at 0.5 + 0.5·(i+1)/N

**Model input:** plain 3-channel RGB per point. No semantic feature encoding.

**Adapted from [jastermark/S23DR2026](https://huggingface.co/jastermark/S23DR2026):**
- Gestalt-guided point sampling and COLMAP projection pipeline
- Post-processing (confidence filtering, vertex merging, gap filling)

## Results

| Approach | Split | F1 | IoU | HSS |
|---|---|---|---|---|
| Perceiver baseline | cleaned val | — | — | 0.350 |
| PointNet two-stage (Path B) | public test | 0.497 | 0.409 | 0.442 |
| WireframeDETR (ours) | cleaned val | 0.603 | 0.471 | 0.534 |
| WireframeDETR (ours, best) | public test | **0.664** | **0.516** | **0.575** |

## Architecture

- Embedding dim: 384, Queries: 128, Encoder layers: 4, Decoder layers: 5
- ~22.7M parameters
- CDN groups: 5, λ_pos=0.4, λ_neg=0.8
- Training: AdamW lr=1e-4, OneCycle schedule, batch=14, 200 epochs, A100 80GB (~27h)

## Checkpoint

`wireframe_detr_cdn_multiscale_384d_128q.pth` — plain RGB, feature_dim=3

## Inference

```python
from s23dr_2026.model import get_model, load_checkpoint_compat
from s23dr_2026.inference import predict_wireframe_v2

import torch
ckpt = torch.load("wireframe_detr_cdn_multiscale_384d_128q.pth", map_location="cpu")
model = get_model(ckpt)
load_checkpoint_compat(model, ckpt)
model.eval().to("cuda")

verts, edges = predict_wireframe_v2(scene, model, "cuda")
```

## Training

```bash
# via Modal
modal run pipeline.py --step train --name my-run

# local
python -m s23dr_2026.train \
  --name my-run \
  --ply_dir /path/to/ply_data \
  --embed_dim 384 --num_queries 128 \
  --num_encoder_layers 4 --num_decoder_layers 5 \
  --use_cdn --cdn_groups 5 \
  --scheduler onecycle --num_epochs 200 \
  --batch_size 14 --device cuda
```

## Credits

- [DN-DETR](https://arxiv.org/abs/2203.01305) — contrastive denoising training
- [S23DR 2026 organisers](https://huggingface.co/spaces/usm3d/S23DR2026) — challenge and baseline
- [jastermark/S23DR2026](https://huggingface.co/jastermark/S23DR2026) — COLMAP projection pipeline and post-processing
- [Modal Labs](https://modal.com) — GPU compute