SEALECT: Scene-Conditioned Pose Priors for Plausible Maritime Vessel Retrieval
Organizations: University of Maryland, Baltimore County · Johns Hopkins University Applied Physics Laboratory
Abstract
We address scene-conditioned object retrieval for compositional augmentation in maritime imagery, where strong geometric constraints make appearance-only retrieval produce physically implausible vessel insertions. We propose SEALECT, a retrieval framework that jointly models visual compatibility, vessel pose, scale, and scene geometry through a unified score combining neural similarity, scene-conditioned pose priors, and observed-pose consistency. We construct a maritime scene-pose library and model the pose and scale priors nonparametrically, as a kernel that conditions each prior continuously on the scene with no learned parameters. Across seven maritime benchmarks, our trained scene-conditioned retriever exceeds the strongest appearance encoder (SigLIP 2) in recall (+0.14), and the training-free prior raises geometric plausibility further at negligible recall cost, more than halving the constraint violations of appearance retrieval and outperforming learned foreground-search baselines (GALA-style, DiscoFOS). In a blind study of end-to-end composites (BootPlace placement, DiffHarmony harmonization), SEALECT ranks first among six methods, ahead of random and GALA-style selection and statistically tied with CLIP, DiscoFOS, and DINO.
Figures & tables
| Methods | Recall | NDCG | PS | Recall | NDCG | PS |
|---|---|---|---|---|---|---|
| Random | 0.106 | 0.726 | 1.364 | 0.347 | 0.719 | 1.365 |
| CLIP | 0.432 | 0.801 | 1.903 | 0.645 | 0.782 | 1.731 |
| DINO | 0.445 | 0.810 | 1.947 | 0.637 | 0.791 | 1.776 |
| SigLIP 2 [ 52 ] | 0.493 | 0.814 | 2.042 | 0.671 | 0.799 | 1.843 |
| DiscoFOS [ 63 ] (zero-shot) | 0.071 | 0.786 | 1.414 | 0.164 | 0.789 | 1.459 |
| Method | JS | Gap | JS | Gap | ||
|---|---|---|---|---|---|---|
| Random | 0.241 | 0.969 | 3.72 | 0.238 | 1.016 | 3.78 |
| DINO | 0.159 | 0.531 | 3.29 | 0.166 | 0.639 | 3.41 |
| DiscoFOS (ft) | 0.196 | 0.534 | 3.21 | 0.186 | 0.616 | 3.37 |
| GALA-style | 0.444 | 0.923 | 1.99 | 0.354 | 0.667 | 2.39 |
| only † | 0.548 | 2.659 | 0.12 | 0.539 | 2.625 | 0.15 |
| Method | LL | Viol. | LL | Viol. |
|---|---|---|---|---|
| Random | -3.053 | .120 | -3.054 | .119 |
| DINO | -2.699 | .066 | -2.751 | .074 |
| DiscoFOS (ft) | -2.790 | .075 | -2.806 | .077 |
| GALA-style | -2.861 | .083 | -2.817 | .081 |
| only † | -1.963 | .006 | -1.968 | .006 |
| Geometry term | Recall | NDCG | PS | Recall | NDCG | PS |
|---|---|---|---|---|---|---|
| only | 0.634 | 0.830 | 2.278 | 0.809 | 0.814 | 2.066 |
| only (geom. A) † | 0.028 | 0.940 | 1.812 | 0.039 | 0.935 | 1.825 |
| GALA-style only [ 72 ] (geom. B) | 0.107 | 0.813 | 1.672 | 0.194 | 0.812 | 1.695 |
| GALA-style (learned) | 0.588 | 0.816 | 2.227 | 0.775 | 0.805 | 2.047 |
| ( Sealect ) | 0.629 | 0.846 | 2.408 | 0.800 | 0.830 | 2.201 |
| Method | BT | CI | Sealect pref. |
|---|---|---|---|
| Sealect | 0.213 | – | |
| CLIP | |||
| DiscoFOS | |||
| DINO | |||
| Random | |||
| GALA-style |
| Yaw | Joint Pose | Y-bin | Scale | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Method | JS | Gap | JS | Gap | JS | Gap | JS | Gap | |||||
| Random | 20 | 0.0812 | 0.4172 | 1.9089 | 0.2410 | 0.9688 | 3.7151 | 0.2258 | 0.5917 | 2.3623 | 0.1612 | 0.5103 | 2.5549 |
| Random | 100 | 0.0798 | 0.4174 | 1.9136 | 0.2377 | 1.0157 | 3.7751 | 0.2239 | 0.5938 | 2.3654 | 0.1615 | 0.5090 | 2.5582 |
| CLIP | 20 | 0.0521 | 0.2570 | 1.7321 | 0.1703 | 0.5932 | 3.3490 | 0.1730 | 0.5420 | 2.3622 | 0.0730 | 0.2991 | 2.3239 |
| CLIP | 100 | 0.0556 | 0.2833 | 1.7797 | 0.1797 | 0.6989 | 3.4654 | 0.1869 | 0.5631 | 2.3875 | 0.0880 | 0.3364 | 2.3475 |
| DINO | 20 | 0.0508 | 0.2312 | 1.7065 | 0.1591 | 0.5307 | 3.2935 | 0.1707 | 0.5001 | 2.3201 | 0.0766 | 0.3183 | 2.3303 |
| Log-Likelihood | Violation Rate | ||||||||
| Method | Yaw | Joint | Y-bin | Scale | Yaw | Joint | Y-bin | Scale | |
| Random | 20 | -1.477 | -3.053 | -2.188 | -2.197 | .024 | .120 | .060 | .028 |
| Random | 100 | -1.475 | -3.054 | -2.183 | -2.197 | .024 | .119 | .060 | .027 |
| CLIP | 20 | -1.290 | -2.747 | -1.983 | -1.739 | .012 | .073 | .033 | .010 |
| CLIP | 100 | -1.330 | -2.813 | -2.043 | -1.806 | .014 | .081 | .037 | .013 |
| DINO | 20 | -1.274 | -2.699 | -1.948 | -1.762 | .013 | .066 | .029 | .010 |
| Scene prior | Recall | NDCG | PS | Recall | NDCG | PS |
|---|---|---|---|---|---|---|
| Discrete bin (ablation) | 0.598 | 0.843 | 2.354 | 0.767 | 0.827 | 2.142 |
| Kernel, | 0.634 | 0.846 | 2.407 | 0.805 | 0.830 | 2.199 |
| Kernel, (default) | 0.627 | 0.848 | 2.418 | 0.801 | 0.832 | 2.212 |
| Kernel, | 0.596 | 0.852 | 2.420 | 0.774 | 0.834 | 2.214 |
| Recall | NDCG | PS | Recall | NDCG | PS | |
|---|---|---|---|---|---|---|
| (default) | 0.627 | 0.848 | 2.418 | 0.801 | 0.832 | 2.212 |
| 0.637 | 0.843 | 2.384 | 0.813 | 0.827 | 2.170 | |
| 0.600 | 0.853 | 2.453 | 0.763 | 0.838 | 2.261 | |
| 0.629 | 0.846 | 2.408 | 0.800 | 0.830 | 2.201 | |
| 0.623 | 0.850 | 2.430 | 0.797 | 0.834 | 2.224 | |
| Geo | Euler MAE (deg) | |||||
|---|---|---|---|---|---|---|
| Dataset | mean | med | Pitch | Yaw † | Roll | |
| ABOShips | 25 | 13.9 | 9.3 | 2.8 | 13.4 | 1.3 |
| IH-UAV | 25 | 13.2 | 9.7 | 5.0 | 7.4 | 7.6 |
| KOLOMVERSE | 25 | 27.2 | 26.6 | 6.3 | 24.3 | 5.0 |
| LaRS | 25 | 28.2 | 23.2 | 7.9 | 22.2 | 7.2 |
| MARVEL | 25 | 15.4 | 10.5 | 3.8 | 12.9 | 2.7 |
| Method | Compat. | Mean Rank | Top-1 | |
|---|---|---|---|---|
| Random | 300 | 0.285 | 5.68 | 6.3% |
| CLIP | 300 | 0.345 | 4.64 | 11.7% |
| DINO | 300 | 0.370 | 4.57 | 8.0% |
| DiscoFOS | 213 | 0.477 | 4.69 | 10.3% |
| GALA-style | 210 | 0.517 | 3.37 | 10.5% |
| only † | 300 | 0.429 | 4.28 | 3.3% |
| Condition | AP | FP / empty-img | |
|---|---|---|---|
| No augmentation | |||
| Plausible geom. | |||
| Implausible geom. |
| ESS med. | ESS 10–90% | ESS | Top-10 | |
|---|---|---|---|---|
| 0.5 | 2.3 | 1.0–18.7 | 73% | 1.00 |
| 1.0 | 14.8 | 1.9–507 | 30% | 0.68 |
| 2.0 | 366 | 13.0–8,442 | 4% | 0.09 |