Glenn Jocher
Yonghye Kwon
guigarfr
perry0418
Josh Veitch-Michaelis
Ttayu
Daniel Suess
Fatih Baltacı
Gabriel Bianconi
IlyaOvodov
Marc
e96031413
Chang Lee
Dustin Kendall
Falak
Francisco Reveriano
FuLin
GoogleWiki
Jason Nataprawira
Jeremy Hu
LinCoce
LukeAI
NanoCode012
NirZarrabi
Oulbacha Reda
Piotr Skalski
SergioSanchezMontesUAM
Shiwei Song
Thomas Havlik
Timothy M. Shead
2021-01-13
<p>This release is a minor update implementing numerous bug fixes, feature additions and performance improvements from <a href="https://github.com/ultralytics/yolov5">https://github.com/ultralytics/yolov5</a> to this repo. Models remain unchanged from v9.0 release.</p>
Branch Notice
<p>The <a href="https://github.com/ultralytics/yolov3">ultralytics/yolov3</a> repository is now divided into two branches:</p>
<ul>
<li><a href="https://github.com/ultralytics/yolov3/tree/master">Master branch</a>: Forward-compatible with all <a href="https://github.com/ultralytics/yolov5">YOLOv5</a> models and methods (<strong>recommended</strong>).<pre><code class="lang-bash">$ git clone https://github.com/ultralytics/yolov3 # master branch (default)
</code></pre>
</li>
<li><a href="https://github.com/ultralytics/yolov3/tree/archive">Archive branch</a>: Backwards-compatible with original <a href="https://pjreddie.com/darknet/">darknet</a> *.cfg models (⚠️ no longer maintained). <pre><code class="lang-bash">$ git clone -b archive https://github.com/ultralytics/yolov3 # archive branch
</code></pre>
</li>
</ul>
<p><img src="https://user-images.githubusercontent.com/26833433/100382066-c8bc5200-301a-11eb-907b-799a0301595e.png" width="800"></p>
<p>** GPU Speed measures end-to-end time per image averaged over 5000 COCO val2017 images using a V100 GPU with batch size 32, and includes image preprocessing, PyTorch FP16 inference, postprocessing and NMS. EfficientDet data from <a href="https://github.com/google/automl">google/automl</a> at batch size 8.</p>
Pretrained Checkpoints
Model
AP<sup>val</sup>
AP<sup>test</sup>
AP<sub>50</sub>
Speed<sub>GPU</sub>
FPS<sub>GPU</sub>
params
FLOPS
<a href="https://github.com/ultralytics/yolov3/releases">YOLOv3</a>
43.3
43.3
63.0
4.8ms
208
61.9M
156.4B
<a href="https://github.com/ultralytics/yolov3/releases">YOLOv3-SPP</a>
<strong>44.3</strong>
<strong>44.3</strong>
<strong>64.6</strong>
4.9ms
204
63.0M
157.0B
<a href="https://github.com/ultralytics/yolov3/releases">YOLOv3-tiny</a>
17.6
34.9
34.9
<strong>1.7ms</strong>
<strong>588</strong>
8.9M
13.3B
<p><strong> AP<sup>test</sup> denotes COCO <a href="http://cocodataset.org/#upload">test-dev2017</a> server results, all other AP results denote val2017 accuracy. </strong> All AP numbers are for single-model single-scale without ensemble or TTA. <strong>Reproduce mAP</strong> by <code>python test.py --data coco.yaml --img 640 --conf 0.001 --iou 0.65</code><br>
<strong> Speed<sub>GPU</sub> averaged over 5000 COCO val2017 images using a GCP <a href="https://cloud.google.com/compute/docs/machine-types#n1_standard_machine_types">n1-standard-16</a> V100 instance, and includes image preprocessing, FP16 inference, postprocessing and NMS. NMS is 1-2ms/img. </strong>Reproduce speed<strong> by <code>python test.py --data coco.yaml --img 640 --conf 0.25 --iou 0.45</code> </strong> All checkpoints are trained to 300 epochs with default settings and hyperparameters (no autoaugmentation).
<strong> Test Time Augmentation (<a href="https://github.com/ultralytics/yolov5/issues/303">TTA</a>) runs at 3 image sizes. </strong>Reproduce TTA** by <code>python test.py --data coco.yaml --img 832 --iou 0.65 --augment</code></p>
Requirements
<p>Python 3.8 or later with all <a href="https://github.com/ultralytics/yolov3/blob/master/requirements.txt">requirements.txt</a> dependencies installed, including <code>torch>=1.7</code>. To install run:</p>
<pre><code class="lang-bash">$ pip install -r requirements.txt
</code></pre>
https://doi.org/10.5281/zenodo.4435632
oai:zenodo.org:4435632
Zenodo
https://github.com/ultralytics/yolov3/tree/v9.1
https://doi.org/10.5281/zenodo.2624707
info:eu-repo/semantics/openAccess
Other (Open)
ultralytics/yolov3: v9.1 - YOLOv5 Forward Compatibility Updates
info:eu-repo/semantics/other