SAR-VesselBench: 32 vessel-detection checkpoints
The 32 fine-tuned dark-vessel detectors reported in Label-Efficient Dark-Vessel Detection in SAR: Does SAR-Domain Pretraining Outperform Optical and ImageNet Transfer across ViT and CNN? (AIPR 2026). Code and result evidence: github.com/jroth1414/SAR-VesselBench.
Two backbone tracks (ViT-B/16 and ConvNeXt-V2-Base) each compare four encoder initializations (random, optical remote sensing, SAR, and ImageNet) at four nested training budgets (12, 28, 56 and 111 xView3-SAR scenes). Every model shares one point-based heatmap detector, optimizer, schedule, scene split and seed (0), and was trained in strict FP32 on H100 GPUs.
Each file is one model's best development checkpoint: a PyTorch Lightning
checkpoint with the full detector (encoder, input adapter and 256-channel
head; about 90M trainable parameters in the ViT track and 94M in the CNN
track). TRAINING_COHORT.json binds every checkpoint's SHA-256, its selected
epoch and its development-selected detection threshold. Verify any download
against it.
Results
F1 on the 16-scene test split / the 50 human-verified scenes, each at the model's development-selected threshold (seed 0):
| Track | Init | 12 scenes | 28 scenes | 56 scenes | 111 scenes |
|---|---|---|---|---|---|
| ViT-B/16 | Random | 0.693 / 0.412 | 0.773 / 0.456 | 0.809 / 0.501 | 0.831 / 0.522 |
| ViT-B/16 | Optical (SatDINO) | 0.806 / 0.486 | 0.837 / 0.506 | 0.837 / 0.515 | 0.857 / 0.562 |
| ViT-B/16 | SAR (SARMAE) | 0.804 / 0.493 | 0.772 / 0.425 | 0.865 / 0.568 | 0.886 / 0.559 |
| ViT-B/16 | ImageNet (AugReg) | 0.813 / 0.523 | 0.848 / 0.500 | 0.859 / 0.551 | 0.851 / 0.558 |
| ConvNeXt-V2-B | Random | 0.657 / 0.409 | 0.706 / 0.441 | 0.780 / 0.492 | 0.816 / 0.511 |
| ConvNeXt-V2-B | Optical (BigEarthNet-v2 S2) | 0.720 / 0.457 | 0.751 / 0.479 | 0.821 / 0.534 | 0.840 / 0.552 |
| ConvNeXt-V2-B | SAR (BigEarthNet-v2 S1) | 0.800 / 0.501 | 0.803 / 0.493 | 0.848 / 0.528 | 0.822 / 0.478 |
| ConvNeXt-V2-B | ImageNet (FCMAE + sup.) | 0.792 / 0.464 | 0.846 / 0.500 | 0.873 / 0.562 | 0.888 / 0.574 |
Main findings (see the paper for intervals and caveats):
- With 12 training scenes, all six pretrained encoders beat random initialization on both held-out sets; all 12 paired scene-bootstrap 95% intervals exclude zero.
- SAR pretraining does not consistently beat optical or ImageNet pretraining: SAR wins 6 of 16 paired comparisons against optical (optical wins 2) and 3 of 16 against ImageNet (ImageNet wins 7).
- Every model misses nearly all vessels within 2 km of shore (near-shore recall at most 0.021 on the verified set).
All results come from a single training seed.
Files
<arm>-f<budget>-s0/checkpoints/best.ckpt, where <arm> is vitrand,
satdino, sarmae, vitin1k (ViT track) or cnnrand, beS2, beS1,
cnnin1k (CNN track), and <budget> is 10, 25, 50 or 100 (percent of the 111
training scenes: 12, 28, 56, 111). One exception: beS2-f10-s0 finished its
last development evaluation in a separate resumed job, so its selected
checkpoint is beS2-f10-s0/checkpoints/terminal-best.ckpt.
TRAINING_COHORT.json gives each model's exact file path and SHA-256.
An earlier, superseded training run of the same grid (code 1a82d508) stays
available at the tag
august-1a82d508.
It is not the run reported in the paper.
Loading
Use the code from the GitHub repository. The checkpoints store their
hyperparameters with a PosixPath, so load them with the weights-only loader
and rebuild the module:
import pathlib
import torch
from src.train.lit_modules import HeatmapLitModule
with torch.serialization.safe_globals([(pathlib.PurePosixPath, "pathlib.PosixPath")]):
payload = torch.load("cnnin1k-f100-s0/checkpoints/best.ckpt", map_location="cpu", weights_only=True)
module = HeatmapLitModule(init_name=payload["hyper_parameters"]["init_name"], load_weights=False)
module.load_state_dict(payload["state_dict"], strict=True)
module.eval()
Inputs are three-channel decibel images [VH, VV, VH - VV], normalized with
the training statistics in the repository's data/stats.json. Whole-scene
tiled inference is src/eval/infer_scene.py; each model's detection threshold
is in TRAINING_COHORT.json.
License β read before use
Each checkpoint is a fine-tuned derivative of a released pretrained encoder and inherits its terms:
| Checkpoints | Source encoder | Terms |
|---|---|---|
sarmae-* |
SARMAE (SAR-1M) | CC BY-NC 4.0, non-commercial only |
cnnin1k-* |
ConvNeXt-V2-Base FCMAE + ImageNet-1K (timm) | CC BY-NC 4.0, non-commercial only |
satdino-* |
SatDINO ViT-B/16 (fMoW-RGB) | Apache-2.0 |
vitin1k-* |
ViT-B/16 AugReg ImageNet-1K (timm) | Apache-2.0 |
beS1-*, beS2-* |
BigEarthNet-v2 ConvNeXt-V2-Base | MIT (upstream license tag) |
vitrand-*, cnnrand-* |
none (random initialization) | MIT (this project) |
The fine-tuning data is xView3-SAR (Paolo et al., NeurIPS 2022), CC BY-NC-SA 4.0; its imagery and labels keep their own terms.
Citation
John Roth, Kyle Wagner, and Liv d'Aliberti. Label-Efficient Dark-Vessel Detection in SAR: Does SAR-Domain Pretraining Outperform Optical and ImageNet Transfer across ViT and CNN? 54th Applied Imagery Pattern Recognition Workshop (AIPR 2026), Lecture Notes in Computer Science, Springer (to appear).
@inproceedings{roth2026labelefficient,
author = {Roth, John and Wagner, Kyle and d'Aliberti, Liv},
title = {Label-Efficient Dark-Vessel Detection in {SAR}: Does {SAR}-Domain
Pretraining Outperform Optical and {ImageNet} Transfer across
{ViT} and {CNN}?},
booktitle = {54th Applied Imagery Pattern Recognition Workshop (AIPR 2026)},
series = {Lecture Notes in Computer Science},
publisher = {Springer},
year = {2026},
note = {To appear}
}