This repo implements yolov3 and yolact with darknet53 backbone from scratch with pytorch.
| Model | Dataset | mAP@0.5 | Model Checkpoint |
|---|---|---|---|
| YOLOv3_416x416 | PASCAL VOC2012 | 0.74 | Download |
| YOLOv3_416x416 | MS COCO2014 | 0.47 | Download |
| Maskv3_416x416 | MS COCO2014 | 0.51 | Download |
$ cd download
$ wget -O voc2012.zip https://www.dropbox.com/s/wpo5ht4rnphsn8k/voc2012.tar.gz?dl=1
$ wget -O coco2014.zip https://www.dropbox.com/s/ct43kswckwpafpq/coco2014.tar.gz?dl=1
$ tar -xzvf voc2012.zip && rm voc2012.zip
$ tar -xzvf coco2014.zip && rm coco2014.zipDownload the dataset and unzip them under download directory with the following structure.
$ tree -L download
download
├── COCO
│ ├── images
│ ├── labels
│ ├── masks
│ ├── test.csv
│ └── train.csv
└── PASCAL_VOC
├── 100examples.csv
├── 1examples.csv
├── 2examples.csv
├── 8examples.csv
├── images
├── labels
├── test.csv
└── train.csv
6 directories, 8 files$ pip install -r requrements.txt
$ python main.py --config config/yolov3_voc.yml
$ python main.py --config config/yolov3_coco.ymlBy feeding an image into YOLOv3, it will generate dense prediction of objects for each scale. For each cell in each scale, there are three anchors responsible for detecting objects, and each anchor will predict the (x_raw, y_raw, w_raw, h_raw, prob_raw, classes_raw).
the output of YOLOv3 should be further processed as following:
x_offset => sigmoid(x_raw)
y_offset => sigmoid(y_raw)
w_cell => anchor_w * torch.exp(w_raw)
h_cell => anchor_h * torch.exp(h_raw)
prob => sigmoid(prob_raw)
class => torch.argmax(classes_raw)[0]
from model.yolov3 import YOLOv3
in_channels = 3
num_classes = 20
model = YOLOv3(in_channels=in_channels,
num_classes=num_classes)
imgs = torch.randn((1, 3, 416, 416))
outs = model(imgs)
print("scale(13):", outs[0].shape) # (1, 3, 13, 13, 25)
print("scale(26):", outs[1].shape) # (1, 3, 26, 26, 25)
print("scale(52):", outs[2].shape) # (1, 3, 52, 52, 25)The groundtruth label for each image is the position information of each object in the image. Specifically, there are three kinds of target labels representing ground truth object position in each scale. the label for each grid cell is (x_offset, y_offset, w_cell, h_cell, prob, class).
from data.dataset import YOLODataset
from data.transform import get_yolo_transform
# hyperparameters
csv_path = 'download/pascal_voc/test.csv'
img_dir = 'download/pascal_voc/images/'
label_dir = 'download/pascal_voc/labels/'
img_size = 416
scales = [13, 26, 52]
anchors = [
[(0.28, 0.22), (0.38, 0.48), (0.9, 0.78)], # scale 13
[(0.07, 0.15), (0.15, 0.11), (0.14, 0.29)], # scale 26
[(0.02, 0.03), (0.04, 0.07), (0.08, 0.06)], # scale 52
] # (3, 3, 2)
# load dataset
transform = get_yolo_transform(img_size=img_size, mode='test')
dataset = YOLODataset(csv_file=csv_path,
img_dir=img_dir,
label_dir=label_dir,
anchors=anchors,
transform=transform)
# peek one sample
img, targets = dataset[0]
print("img shape:", img.shape) # (3, 416, 416)
print("number of targets:", len(targets)) # 3
print("t1 shape:", targets[0].shape) # (3, 13, 13, 6)
print("t2 shape:", targets[1].shape) # (3, 26, 26, 6)
print("t3 shape:", targets[2].shape) # (3, 52, 52, 6)