Model Zoo¶
All MegaDetector model variants with performance metrics, parameter counts, and licenses.
MegaDetector V6 (Current)¶
The latest release focuses on efficiency, modern architectures, and licensing flexibility, SMALLER, FASTER, BETTER.
Highlights¶
- 50x smaller: The compact YOLOv10 variant has 2.3M parameters, 2% of MegaDetectorV5's 139.9M, with comparable accuracy
- Multiple architectures: YOLOv9, YOLOv10, and RT-DETR for different hardware budgets
- Permissive licenses: MIT and Apache-2.0 options alongside AGPL-3.0
- Ongoing fine-tuning: V6 weights are refreshed as new public and private data arrives
Model Variants¶
| Model | Params | Animal Recall | mAP50 | License |
|---|---|---|---|---|
| MDV6-apa-rtdetr-e | 76M | 82.9% | 94.1% | Apache-2.0 |
| MDV6-yolov10-e | 29.5M | 82.8% | 92.8% | AGPL-3.0 |
| MDV6-yolov9-e | 58.1M | 82.1% | 88.6% | AGPL-3.0 |
| MDV6-rtdetr-c | 31.9M | 81.6% | 89.9% | AGPL-3.0 |
| MDV6-apa-rtdetr-c | 20M | 81.1% | 91.0% | Apache-2.0 |
| MDV6-yolov9-c | 25.5M | 78.4% | 87.9% | AGPL-3.0 |
| MDV6-yolov10-c | 2.3M | 76.8% | 87.2% | AGPL-3.0 |
| MDV6-mit-yolov9-e | 51M | 76.1% | 71.5% | MIT |
| MDV6-mit-yolov9-c | 9.7M | 74.8% | 87.6% | MIT |
Model names are standardized into MDV6-Compact and MDV6-Extra for the two model sizes within each architecture.
Which Model Should I Use?¶
- Best accuracy: MDV6-apa-rtdetr-e (82.9% recall, Apache-2.0)
- Best for laptops/edge: MDV6-yolov10-c (2.3M params, runs on CPU)
- Best balance: MDV6-yolov10-e (29.5M params, 82.8% recall)
- Need MIT license?: MDV6-mit-yolov9-c
from PytorchWildlife.models import detection as pw_detection
# Load a specific variant
model = pw_detection.MegaDetectorV6(version="MDV6-apa-rtdetr-e")
YOLOv10 variants (MDV6-yolov10-*)¶
YOLOv10 covers the broadest range in the zoo. The extra-size MDV6-yolov10-e (29.5M parameters) reaches 82.8% animal recall, while the compact MDV6-yolov10-c shrinks to 2.3M parameters for laptops and edge devices. Both ship under AGPL-3.0 and are selectable from the megadetector CLI.
RT-DETR variants (MDV6-rtdetr-*)¶
RT-DETR is where the top accuracy sits. MDV6-apa-rtdetr-e records the highest animal recall in the table (82.9%) under an Apache-2.0 license, which suits projects that need permissive terms. A smaller MDV6-apa-rtdetr-c and an AGPL MDV6-rtdetr-c round out the family.
Model Licensing¶
V6 deliberately offers variants under three licenses so you can match the model to your project's distribution requirements:
| License | Variants | Use when |
|---|---|---|
| MIT | MDV6-mit-yolov9-c, MDV6-mit-yolov9-e |
You need a permissive license with no copyleft obligations, e.g. bundling weights into a closed-source product |
| Apache-2.0 | MDV6-apa-rtdetr-c, MDV6-apa-rtdetr-e |
You want a permissive license with an explicit patent grant; MDV6-apa-rtdetr-e is also the top-accuracy variant |
| AGPL-3.0 | the remaining YOLOv9/YOLOv10/RT-DETR variants | Your use is compatible with strong copyleft (research, internal tools, AGPL-licensed services) |
The repository code is MIT-licensed independently of the weights. Always confirm the license of the specific variant you ship.
[!NOTE] The
megadetectorCLI selects from the AGPL YOLOv9/YOLOv10/RT-DETR variants via--model; the MIT and Apache variants are loaded through the PyTorch-Wildlife Python API.
Performance Benchmarks: GPU and Edge CPU Inference¶
| Hardware | Model | Approximate Speed |
|---|---|---|
| NVIDIA RTX 3090 | MDV6-yolov10-c (2.3M) | ~100–200 images/sec |
| NVIDIA RTX 3090 | MDV6-yolov10-e (29.5M) | ~30–60 images/sec |
| Modern CPU (no GPU) | MDV6-yolov10-c (2.3M) | ~2–5 images/sec |
| Google Colab (free GPU) | Any V6 variant | ~10–50 images/sec |
At 50 images/sec on a GPU, one million images takes about 5.5 hours. On CPU with the compact model, about 3.9 days. Every V6 variant is faster than V5 (139.9M params).
Edge and CPU inference: the compact MDV6-yolov10-c (2.3M parameters) is the variant to target for CPU-only and edge hardware. It needs no GPU and runs at roughly 2–5 images/sec on a modern CPU, which is what makes it a fit for field devices like the SPARROW unit.
Version History¶
| Version | Year | Architecture | Params | Notes |
|---|---|---|---|---|
| V6.0 (current) | 2024 | YOLOv9/v10, RT-DETR | 2.3M–76M | Multiple variants, MIT/Apache options |
| V5.0 | 2022 | YOLOv5 | 139.9M | Two sub-versions (5a, 5b) |
| V4.1 | 2020 | Faster R-CNN | , | Added vehicle class |
| V3 | 2019 | Faster R-CNN | , | Added human class |
| V2 | 2018 | Faster R-CNN | , | First public release |
MegaDetector V5 and Earlier¶
For MegaDetectorV5 model weights and earlier versions, see the archive branch of the Biodiversity repository (formerly microsoft/CameraTraps).
MegaDetector V1–V5 were originally developed by Dan Morris at Microsoft. The V5 weights load directly through PyTorch-Wildlife, so existing V5 pipelines continue to work.