Pytorch implementation of RetinaNet object detection as described in Focal Loss for Dense Object Detection by Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He and Piotr Dollár.
This implementation is primarily designed to be easy to read and simple to modify.
Currently, this repo achieves 33.5% mAP at 600px resolution with a Resnet-50 backbone. The published result is 34.0% mAP. The difference is likely due to the use of Adam optimizer instead of SGD with weight decay.
-
Clone this repo
-
Install the required packages:
apt-get install tk-dev python-tk
- Install the python packages:
pip install pandas
pip install pycocotools
pip install opencv-python
pip install requests
The network can be trained using the train.py script. Currently, two dataloaders are available: COCO and CSV. For training on coco, use
python train.py --dataset coco --coco_path ../coco --depth 50
For training using a custom dataset, with annotations in CSV format (see below), use
python train.py --dataset csv --csv_train <path/to/train_annots.csv> --csv_classes <path/to/train/class_list.csv> --csv_val <path/to/val_annots.csv>
Note that the --csv_val argument is optional, in which case no validation will be performed.
A pre-trained model is available at:
- https://drive.google.com/open?id=1yLmjq3JtXi841yXWBxst0coAgR26MNBS (this is a pytorch state dict)
The state dict model can be loaded using:
retinanet = model.resnet50(num_classes=dataset_train.num_classes(),)
retinanet.load_state_dict(torch.load(PATH_TO_WEIGHTS))
Run coco_validation.py to validate the code on the COCO dataset. With the above model, run:
python coco_validation.py --coco_path ~/path/to/coco --model_path /path/to/model/coco_resnet_50_map_0_335_state_dict.pt
This produces the following results:
Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.335
Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.499
Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.357
Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.167
Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.369
Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.466
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.282
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.429
Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.458
Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.255
Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.508
Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.597
For CSV Datasets (more info on those below), run the following script to validate:
python csv_validation.py --csv_annotations_path path/to/annotations.csv --model_path path/to/model.pt --images_path path/to/images_dir --class_list_path path/to/class_list.csv (optional) iou_threshold iou_thres (0<iou_thresh<1)
It produces following resullts:
label_1 : (label_1_mAP)
Precision : ...
Recall: ...
label_2 : (label_2_mAP)
Precision : ...
Recall: ...
You can also configure csv_eval.py script to save the precision-recall curve on disk.
To visualize the network detection, use visualize.py:
python visualize.py --dataset coco --coco_path ../coco --model <path/to/model.pt>
This will visualize bounding boxes on the validation set. To visualise with a CSV dataset, use:
python visualize.py --dataset csv --csv_classes <path/to/train/class_list.csv> --csv_val <path/to/val_annots.csv> --model <path/to/model.pt>
The retinanet model uses a resnet backbone. You can set the depth of the resnet model using the --depth argument. Depth must be one of 18, 34, 50, 101 or 152. Note that deeper models are more accurate but are slower and use more memory.
The CSVGenerator provides an easy way to define your own datasets.
It uses two CSV files: one file containing annotations and one file containing a class name to ID mapping.
The CSV file with annotations should contain one annotation per line. Images with multiple bounding boxes should use one row per bounding box. Note that indexing for pixel values starts at 0. The expected format of each line is:
path/to/image.jpg,x1,y1,x2,y2,class_name
Some images may not contain any labeled objects.
To add these images to the dataset as negative examples,
add an annotation where x1, y1, x2, y2 and class_name are all empty:
path/to/image.jpg,,,,,
A full example:
/data/imgs/img_001.jpg,837,346,981,456,cow
/data/imgs/img_002.jpg,215,312,279,391,cat
/data/imgs/img_002.jpg,22,5,89,84,bird
/data/imgs/img_003.jpg,,,,,
This defines a dataset with 3 images.
img_001.jpg contains a cow.
img_002.jpg contains a cat and a bird.
img_003.jpg contains no interesting objects/animals.
The class name to ID mapping file should contain one mapping per line. Each line should use the following format:
class_name,id
Indexing for classes starts at 0. Do not include a background class as it is implicit.
For example:
cow,0
cat,1
bird,2
- Significant amounts of code are borrowed from the keras retinanet implementation
- The NMS module used is from the pytorch faster-rcnn implementation
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent