01
Holothurian
Sea cucumber
Reference placeholderMARINE VISION. LIGHTWEIGHT BY DESIGN.
Intelligence.
Below the
surface.
Meet WAYTECHG Marine Organism Detection. A compact deep-learning pipeline built to explore more efficient ways of seeing life on the seafloor.
Built for discovery.AI Engineering · Xiamen University Malaysia
OUR FOCUSFour species.
One ecosystem.
One ecosystem.
FINAL RESEARCH MODEL
4.37M
Parameters↓ 7.02%
Dive into the researchDU-MobileYOLO / GhostConv / SimSPPF
04Target organisms
07Model variants
7.02%Fewer parameters
75.36%Final mAP@0.5
Reported research results
Final method vs. DU-MobileYOLO
01 / THE RESEARCHA lighter way forward
Less complexity.
More possibility.
Underwater scenes are complex. The network doesn't have to be heavy.
Explore the URPC2020 datasetWAYTECHG studies a lightweight deep-learning pipeline for fast object detection of four target aquatic species. The final method combines GhostConv and SimSPPF within DU-MobileYOLO to reduce model size and computation while improving reported mAP@0.5.
A
Academic objective
Measure the accuracy–efficiency tradeoff through a baseline comparison and controlled module ablations on URPC2020.
B
Practical objective
Explore compact detection for resource-limited marine monitoring and future autonomous underwater vehicle workflows.
Make yourself at home in the dashboard. Upload an image to compare the two methods. Check the mode indicator: a local training checkout runs real checkpoints; the standalone public demo uses illustrative boxes. Select chart axes and a model to inspect the measured research metrics.
Four small inhabitants. A shared habitat.
ILLUSTRATED REFERENCE PLACEHOLDERS02
Scallop
Marine bivalve
Reference placeholder03
Echinus
Sea urchin
Reference placeholder04
Starfish
Sea star
Reference placeholderClass illustrations stand in for ground-truth dataset crops. They are not URPC2020 photographs or annotations.
02 / THE LIVE DASHBOARDExplore. Compare. Understand.
Explore the pipeline.
Find your perspective.
Your image. Two methods. Six models to compare.
YOUR IMAGE, TWO PERSPECTIVES
Take the pipeline for a swim.
Simulation mode. Boxes, species labels, and confidence scores are illustrative. This public demo does not execute trained model weights or evaluate the accuracy of your image.
Drop an image. Discover the workflow.
JPG, PNG, or WebP · Up to 8 MB and 20 megapixels
Preview appears above. Press Compare models to show results below. Displays are resized to a filled 640 × 640 square; the original image is used for inference.
Choose your image or try the illustrated scene.
DU-MobileYOLO
BaselineBaseline resultChoose an image above, then press Compare models.
GhostConv + SimSPPF
Final proposedFinal proposed resultChoose an image above, then press Compare models.
Uploads are processed in memory and are not saved by this app.Model metrics are from the offline study.
SIX MODELS. YOUR POINT OF VIEW.
Find the balance that matters.
Highest mAP@0.5 in this comparison75.36%
Final proposed · reported research result
Smallest parameter count4.37M
Ghost + SPPF and final proposed · tied
Fastest reported GPU throughput50.90FPS
Final proposed · RTX 4060 Laptop GPU
THE COMPLETE STUDY
Every change, in context.
Inspect the baseline, four ablation variants, and the final proposed method. Changes below are relative to DU-MobileYOLO; percentage points describe accuracy changes.
Select a point or model below to inspect its metrics.
All model metrics
| Model | mAP@0.5 | mAP@0.5:0.95 | Params (M) | GFLOPs | FPS (GPU) | Latency (ms) | Precision | Recall | Holothurian AP@0.5 | Echinus AP@0.5 | Scallop AP@0.5 | Starfish AP@0.5 |
|---|
Accuracy and complexity: supplied report, page 11, Tables I–III. FPS: sustained profiling logs, runs 5–10, on an RTX 4060 Laptop GPU, including inference and NMS. These are not CPU cloud performance claims. Unreported values remain blank. The final model improves mAP@0.5; mAP@0.5:0.95 is 0.30 percentage points lower than baseline.
03 / KEEP THE CONVERSATION FLOWINGIdeas welcome
Good things start
with a hello.
Research questions, thoughtful feedback, or a project worth exploring together. I'd love to hear from you.
WAYTECHGAI Engineering · Xiamen University Malaysia