ROVERCAM / YOLO26M / BROWSER INFERENCE

Try my rover detector yourself.

I trained this CNN locally from scratch on my own custom Cars4Mars dataset. The final training run took about 35 hours on an RTX 3050 Laptop GPU. This page runs the raw ONNX export directly in your browser.

Download raw ONNX Ultralytics model page GitHub repo

Starting ONNX Runtime and loading the rover model…

Precision87.7%
Recall80.7%
mAP5088.7%
mAP50–9569.5%
TRAINED LOCALLY~35 hours
CUSTOM DATASET15,450 training images
HELD-OUT TEST306 images / 3,123 objects
LOCAL GPURTX 3050 Laptop · 4 GB

MODEL

What I trained

ArchitectureUltralytics YOLO26m
TrainingFrom scratch
Input736 × 736 RGB
Classes8
Fused parameters20,355,620
Compute67.9 GFLOPs
Training batch4
Browser ONNX batchAuto-detected (current: 4)
Max epochs175

Trained locally in WSL2 Ubuntu 22.04 with CUDA + AMP. The public accuracy numbers on this page are from the final held-out test set, not an earlier smoke/validation run.

CLASSES

What it can detect

8 custom classes

FINAL TEST

Per-class performance

306 held-out images
ClassPrecisionRecallmAP50mAP50–95
Hammer91.6%83.7%90.6%59.5%
Tennis ball84.5%87.5%85.1%65.8%
Traffic cone100.0%97.9%99.5%86.9%
Black balloon82.8%73.5%84.2%60.8%
Blue balloon94.1%78.4%92.4%75.6%
Pink balloon80.1%72.1%82.7%66.2%
White balloon81.4%72.1%83.8%66.2%
Yellow balloon87.0%80.2%91.1%74.6%

Traffic cone was the strongest class at 99.5% mAP50. Balloon detection is more sensitive to colour, shadows, lighting and overlap.

LIVE TEST

Drop in an image

ONNX Runtime Web / WASM
Load the model, then choose an image.

Browser speed depends on the device running the page. The Android V21 build in the repo is separate: that version uses the compiled QNN context on the Snapdragon 8 Gen 3 Hexagon HTP/NPU.