--- license: cc-by-nc-4.0 tags: - object-detection - underwater - synthetic-data - domain-randomization - yolo --- # MSc Dissertation — Synthetic Underwater Imagery & Sim-to-Real Detection Supporting data and model checkpoints for an MSc dissertation on synthetic benthic-survey imagery generation and sim-to-real object detection. Benthic scenes are generated with diffusion models (Flux.2-dev, Hunyuan), domain-randomised through a Jerlov water-column physics model, then used to train YOLO detectors evaluated zero-shot against the real [DUO dataset](https://github.com/chongweiliu/DUO). Full pipeline code, configs, and training scripts: **[github.com/rithish007/ImgGen_AI](https://github.com/rithish007/ImgGen_AI)** (MIT license). The descoped Unreal Engine 5.8 simulator arm that preceded this pipeline is archived at [github.com/rithish007/Shallow_Seabed](https://github.com/rithish007/Shallow_Seabed) (MIT license). Classes: `starfish`, `sea_urchin`, `scallop`. ## Contents ``` datasets/ flux2dev/ base/ 1000 images — Flux.2-dev v9 generations, no domain randomization placeholder_dr/ 2000 images — base set + placeholder-DR copies duo_scatter_dr/ 2000 images — base set + DUO-calibrated-scatter DR copies (used to train the best model) duo_dr/ 2000 images — base set + DUO-calibrated DR copies hunyuan/ 892 images — Hunyuan v8 generations (generated only; not yet trained or evaluated — see note below) weights/ flux2dev_base.pt YOLO trained on datasets/flux2dev/base flux2dev_placeholder_dr.pt YOLO trained on datasets/flux2dev/placeholder_dr flux2dev_duo_scatter_dr_BEST.pt YOLO trained on datasets/flux2dev/duo_scatter_dr — best DUO mAP50 so far flux2dev_duo_dr.pt YOLO trained on datasets/flux2dev/duo_dr duo_test_detections/ Predictions of flux2dev_duo_scatter_dr_BEST.pt on the real DUO test set *.jpg images with predicted boxes drawn labels/*.txt raw YOLO-format predicted boxes (class x_center y_center width height conf) datasets/shallowseabed/ DESCOPED -- see "ShallowSeabed" section below. Different classes (coral/kelp/rock/sponge), from the Unreal Engine simulator this project moved away from, kept for the reproducibility record only. base/ 500 images (train+val) -- original UE5.8 render, no domain randomization dr_transformed/ 500 images (train+val), same frames/labels -- a DR-transformed render weights/ shallowseabed_no_aug.pt YOLO trained on datasets/shallowseabed/base, no augmentation shallowseabed_with_aug_BEST.pt YOLO trained on datasets/shallowseabed/base, with augmentation -- best of the 4 shallowseabed_no_aug_dr_trans.pt YOLO trained on datasets/shallowseabed/dr_transformed, no augmentation shallowseabed_with_aug_dr_trans.pt YOLO trained on datasets/shallowseabed/dr_transformed, with augmentation ``` Each `datasets/flux2dev/*` and `datasets/shallowseabed/*` folder is a standard YOLO layout (`images/train`, `images/val`, `labels/train`, `labels/val`, `data.yaml`). ## Results: zero-shot evaluation on real DUO test set All four models are trained only on synthetic data and evaluated with no fine-tuning on real imagery. | Variant | Train images | DUO mAP50 | DUO mAP50-95 | |---|---|---|---| | flux2dev_base | 900 | 10.31% | 4.78% | | flux2dev_placeholder_dr | 1800 | 10.40% | 5.21% | | flux2dev_duo_dr | 1800 | 11.21% | 5.84% | | **flux2dev_duo_scatter_dr (best)** | 1800 | **11.82%** | **6.18%** | Domain randomization calibrated against real DUO water statistics (`duo_dr`, `duo_scatter_dr`) consistently outperforms uncalibrated placeholder randomization, which in turn outperforms no randomization at all. ## ShallowSeabed — descoped Unreal Engine simulator arm **This dataset was descoped and is not part of the reported pipeline.** Before adopting the diffusion-model approach (Flux.2-dev/Hunyuan) documented above, this project built an Unreal Engine 5.8 underwater survey simulator with a custom stencil-based instance annotation pipeline — see [github.com/rithish007/Shallow_Seabed](https://github.com/rithish007/Shallow_Seabed). It produces a **different 4-class set** (`coral`, `kelp`, `rock`, `sponge`), not the 3-class DUO set (`starfish`, `sea_urchin`, `scallop`) used everywhere else in this repo. It's included here as the reproducibility record for that abandoned track, not as usable pretrained models. Two 500-frame image sets from the same underlying survey (same frames, same labels, different render/domain transform), each trained no-augmentation vs. with-augmentation: | Variant | mAP50 (val) | mAP50-95 (val) | |---|---|---| | shallowseabed_no_aug | 13.3% | 4.2% | | **shallowseabed_with_aug (best overall)** | **23.5%** | **9.5%** | | shallowseabed_no_aug_dr_trans | 8.1% | 2.6% | | shallowseabed_with_aug_dr_trans | 14.5% | 5.7% | **Why it was descoped — sim-to-real transfer failed at scale.** All four models were run against ~1955 real (non-simulated) underwater photos as a held-out test set: | Model | Real-photo detection rate (≥1 box, conf ≥ 0.25) | |---|---| | shallowseabed_no_aug | 0/1955 (0.0%) | | shallowseabed_with_aug | 2/1955 (0.1%) | | shallowseabed_no_aug_dr_trans | 26/1955 (1.3%) | | shallowseabed_with_aug_dr_trans | 94/1955 (4.8%) | Even the best case detects something in fewer than 5% of real photos — the domain gap holds at scale despite reasonable in-domain (synthetic validation) accuracy above. This result is the concrete evidence behind the project's pivot to the diffusion-model generation pipeline, which closes this gap far better (see the DUO zero-shot results above). The real photos themselves are not republished here (third-party sourced, not original to this project). ## Known limitations - The Hunyuan dataset was generated but has not yet been used to train or evaluate a model — it's included here for completeness/reproducibility, not as a paired dataset+weights set. - Scallop and diver false-positive behaviour is under active investigation; see the GitHub repo for the current state of that analysis. ## License This dataset (images, labels, and model checkpoints) is released under [CC-BY-NC-4.0](https://creativecommons.org/licenses/by-nc/4.0/) — reuse and adaptation are permitted for non-commercial purposes, with attribution. For commercial licensing, contact the author. The companion `ImgGen_AI` code repository is released separately under the MIT License. ## Citation If you use this dataset, please cite it: ```bibtex @dataset{ramamoorthysathya2026dataset, author = {Ramamoorthy Sathya, Rithish}, title = {{MSc Dissertation Datasets and Model Checkpoints: Synthetic Underwater Imagery and YOLO Detectors}}, year = {2026}, publisher = {Hugging Face}, url = {https://huggingface.co/datasets/Rithish007/MScDissertation}, note = {CC-BY-NC-4.0. University of Sheffield MSc Robotics Dissertation.} } ``` If you use the generation/training pipeline, please also cite the code: ```bibtex @misc{ramamoorthysathya2026imggen, author = {Ramamoorthy Sathya, Rithish}, title = {{ImgGen\_AI: Synthetic Underwater Imagery Generation and Sim-to-Real Object Detection}}, year = {2026}, howpublished = {\url{https://github.com/rithish007/ImgGen_AI}}, note = {Code repository} } ``` If you use the descoped Unreal Engine simulator arm, please cite it separately: ```bibtex @misc{ramamoorthysathya2026shallowseabed, author = {Ramamoorthy Sathya, Rithish}, title = {{ShallowSeabed: Unreal Engine 5 Synthetic Underwater Scene Simulator}}, year = {2026}, howpublished = {\url{https://github.com/rithish007/Shallow_Seabed}}, note = {Code repository} } ```