# Checkpoints

This folder contains the model checkpoints corresponding to the headline
WildBox results reported in the paper, one per architecture.

```
checkpoints/
├── CHECKPOINTS_README.md
├── detany3d_ep2_seed0/
│   └── checkpoint_0.pth                 ← 4.1 GB — DetAny3D headline row
└── ovmono3d_lift_init5sp_seed0/
    └── model_final.pth                  ← 574 MB — OVMono3D-LIFT headline row
```

---

## OVMono3D-LIFT — `ovmono3d_lift_init5sp_seed0/model_final.pth`

* **Architecture:** OVMono3D-LIFT (Cube R-CNN backbone + GLIP open-vocabulary
  2D detector, 3D head fine-tuned on WildBox).
* **Training schedule:** 10 000 iters of 5-species pretraining (wildlife5)
  → 10 000 iters of 6-species WildBox continuation = 20 000 total iters.
* **Reproduces:** the row labelled *"init5sp curriculum, 10 000 iters"* in the
  paper's fine-tuning table — macro 3D AP ≈ 13.17, BEV AP @ 0.50 ≈ 8.68 on
  `WildBox_val_paper.json`.
* **Format:** standard detectron2 / Cube R-CNN state dict.

### How to load

From the OVMono3D-LIFT eval repo:

```bash
python tools/train_net.py \
    --eval-only \
    --config-file configs/wildbox/OVMono3D_wildbox_wildlife6.yaml \
    MODEL.WEIGHTS path/to/checkpoints/ovmono3d_lift_init5sp_seed0/model_final.pth
```

The other two seeds used to compute the ± std bar in that row are not shipped;
they can be reproduced from the OVMono3D-LIFT eval repo with seeds 1 and 2.

---

## DetAny3D — `detany3d_ep2_seed0/checkpoint_0.pth`

* **Architecture:** DetAny3D (SAM-ViT-H backbone + UniDepth depth prior +
  multi-prompt 3D head, with Grounding-DINO oracle 2D for prompts).
* **Training schedule:** 2-epoch fine-tune on WildBox training split, seed 0,
  initialised from the upstream DetAny3D pretrained checkpoint. Note that
  fine-tuning to 3 epochs **overfits** — see the ablation table in the paper.
* **Reproduces:** the row labelled *"DetAny3D fine-tuned (2 epochs, oracle 2D)"*
  in the paper's headline table.
* **Format:** native DetAny3D `checkpoint_<epoch>.pth` (the file is at epoch 1,
  i.e. *after* 2 training epochs — DetAny3D 0-indexes its checkpoint counter).

### How to load

From the DetAny3D eval repo (`tools/wildbox_final.sh` wraps the full eval):

```bash
torchrun --nproc_per_node=4 ./train.py \
    --config_path detect_anything/configs/wildbox/wildbox_eval_oracle2d.yaml \
    --resume path/to/checkpoints/detany3d_ep2_seed0/checkpoint_0.pth \
    --inference-only
```

Other seeds are not shipped; the headline number in the paper is single-seed
(see the paper's ablation table for the seed-1 / seed-2 spread reasoning).

---

**Release note (v1.1).** These are the paper's headline checkpoints, trained and
evaluated against the paper-version annotation export. This repository ships the
v1.1 annotation export (human-corrected tracks; see the dataset README), so
re-evaluating these checkpoints against the v1.1 ground truth can yield numbers
that differ from the paper's, which are reproduced against the paper-version
ground truth.