ObjectDetection.Pytorch
This repo is an object detection library for pytorch (single stage detectors).
RetinaNet Architecture
VOCmAP: train VOC07+12 test VOC07
| backbone | resolution | VOCmAP | COCOmAP | Inference[ms] | model |
|---|---|---|---|---|---|
| VGG16 | 300 | 79.5 | here | ||
| resnet18 | 300 | 76.5 | |||
| resnet50 | 300 | 80.5 | |||
| resnet101 | 300 | ||||
| resnet18 | 600 | ||||
| resnet50 | 600 | ||||
| resnet101 | 600 |
To start off
requirements: cv2, pandas. plz install.
clone the repo.
git clone https://github.com/kentaroy47/ObjectDetection.Pytorch.git
Download PASCALVOC2007 dataset and extract.
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
Download reduced FC vgg weights and place in weights folder.
mkdir weights
cd weights
wget https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth
Train SSD Models
run train_ssd.ipynb
to run inference, try inference.ipynb.
the trained SSD model is here (still underfitting..) https://github.com/kentaroy47/ObjectDetection.Pytorch/releases/download/ssdvgg200/ssd300_200.pth
Train YOLO Models
run yolo.ipynb (TBD)
Train Faster RCNN Models
run frcnn.ipynb (TBD)
Test models
run eval.ipynb
Test results for ssd
Pascal VOC 2007 test set.
SSD-300
Model: https://github.com/kentaroy47/ObjectDetection.Pytorch/releases/download/ssdvgg200/ssd300_200.pth
Mean AP = 0.7959
Results:
0.842
0.850
0.784
0.736
0.518
0.891
0.888
0.902
0.634
0.832
0.793
0.873
0.899
0.862
0.815
0.521
0.798
0.815
0.885
0.780
0.796
RetinaNet-300 Resnet18
Mean AP = 0.7279
Results:
0.759
0.814
0.725
0.661
0.373
0.807
0.836
0.847
0.508
0.759
0.741
0.816
0.848
0.813
0.743
0.420
0.695
0.803
0.859
0.731
0.728