RetinaNet

RetinaNet Object Detection on Satellite Images

Pre-requistes

• Colab
• Keras and Tensorflow
• Clone into https://github.com/fizyr/keras-retinanet.git
• Install requirements by pip install . and Build cython code python setup.py build_ext --inplace

Preprocessing

In first step, separate images into train,valid and test directories. The dataset was provided in Pascal voc but was converted to custom format of filepath,x1,y1,x2,y2,class_name. Training requires annotation files in .CSV format. For step by step preprocessing please refer to Preprocessing/RetinaNet_preprocessing.ipynb

Example dataset provided in Dataset folder

Create classes.csv in format:

class_name,id
car,0
bus,1
...

Network Summary

RetinaNet (ResNet50 backnone)

RetinaNet (EfficientNetB7 backnone)

Debug Annotations

Check validity of annotations by running following commands:

!python3 keras_retinanet/bin/debug.py csv labels_train.csv sims_classes.csv  

Training

Form more details please refer to '*.ipynb` availble in current directory.
For training on custom dataset, a CSV file can be used to pass data. To train using CSV run:

!python3 keras_retinanet/bin/train.py --batch-size 4 --epochs 50 --steps 1000 --tensorboard-dir '/content/drive/My Drive/SIMS_Dataset/baseline/runs/' 
--snapshot-path '/content/drive/My Drive/SIMS_Dataset/baseline/snapshot' --no-resize --weighted-average --compute-val-loss 
csv labels_train.csv sims_classes.csv --val-annotations labels_valid.csv

Testing

Convert Trained model to inference model by this command:

!python3 keras_retinanet/bin/convert_model.py '/content/drive/My Drive/SIMS_Dataset/baseline/snapshot/resnet50_csv_20.h5' snapshots-inference/retinaNet_resnet50.h5

Form more details please refer to '*.ipynb` availble in current directory.
Run:

python3 keras_retinanet/bin/evaluate.py --save-path test_output csv labels_test.csv sims_classes.csv 'snapshots-inference/retinaNet_resnet50.h5'

Performance Measures

Model	Validation mAP	Test mAP
RetinaNet (ResNet50)	0.8442	0.6231
RetinaNet (EfficientNetB7)	0.6126	evaluation script error-see this issue

Performance Graphs

Graphs for baseline (ResNet50)

Training settings

Image Size = 1024x768

Batch_size = 4

iterations_per_epoch = 1000

LR = 1e-5

Graphs for improvement with EfficientNetB7 as Backbone

Training settings

Image Size = 512

Batch_size = 1

iterations_per_epoch = 3000

LR = 1e-5

Per Class Evaluation results

Validation set (ResNet50)

Running network: 100% (1740 of 1740) |####| Elapsed Time: 0:02:51 Time:  0:02:51
Parsing annotations: 100% (1740 of 1740) || Elapsed Time: 0:00:00 Time:  0:00:00
11462 instances of class car with average precision: 0.8828
864 instances of class truck with average precision: 0.6061
2198 instances of class van with average precision: 0.7179
381 instances of class longvehicle with average precision: 0.5732
631 instances of class bus with average precision: 0.7181
212 instances of class airliner with average precision: 0.9720
41 instances of class propeller with average precision: 0.5170
141 instances of class trainer with average precision: 0.9471
155 instances of class chartered with average precision: 0.8477
8 instances of class fighter with average precision: 0.7901
190 instances of class other with average precision: 0.1471
126 instances of class stairtruck with average precision: 0.2763
68 instances of class pushbacktruck with average precision: 0.1446
16 instances of class helicopter with average precision: 0.1049
4707 instances of class boat with average precision: 0.9416
mAP: 0.8442

Test set (ResNet50)

3678 instances of class car with average precision: 0.8669
446 instances of class truck with average precision: 0.6585
874 instances of class van with average precision: 0.7033
222 instances of class longvehicle with average precision: 0.5817
366 instances of class bus with average precision: 0.7277
161 instances of class airliner with average precision: 0.9638
25 instances of class propeller with average precision: 0.6700
49 instances of class trainer with average precision: 0.8611
103 instances of class chartered with average precision: 0.8428
8 instances of class fighter with average precision: 0.8409
85 instances of class other with average precision: 0.1591
83 instances of class stairtruck with average precision: 0.2610
55 instances of class pushbacktruck with average precision: 0.1585
9 instances of class helicopter with average precision: 0.1115
1820 instances of class boat with average precision: 0.9398
Inference time for 747 images: 0.0855
mAP using the weighted average of precisions among classes: 0.8211
mAP: 0.6231

Visual Results