Yolo V3 Tensorflow Implementation:

In this repo, I have re-implementated famous and revolutionary object detection algorithm YoloV3.

Get Started:

1. Prepare dataset

• Download and extract VOC datasets from following links. Combine 2007 and 2012 train+val set in single directory & 2007 and 2012 test set in single directory.

• Training Set: VOC-2007 train+val : http://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar VOC-2012 train+val : http://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar

• Testing Set: VOC-2007 test : http://pjreddie.com/media/files/VOCtest_06-Nov-2007.tar VOC-2012 test : http://pjreddie.com/media/files/VOC2012test.tar

2. [OPTIONAL] Convert voc dataset into csv.

• In case you have different dataset, then convert the dataset images and labels into a csv file having following format. img_path,img_h,img_w,bbox_1_xmin,bbox_1_ymin,bbox_1_xmax,bbox_1_ymax,bbox_1_cls_id,bbox_2_xmin,bbox_2_ymin,bbox_2_xmax,bbox_2_ymax,bbox_2_cls_id,...,bbox_N_xmin,bbox_N_ymin,bbox_N_xmax,bbox_N_ymax,bbox_N_cls_id

• For VOC dataset, execute following command.

• For train set csv preparation: python ./misc_files/voc_to_csv_conversion.py --dataset_path "./PASCAL/Combined/" --dataset Train --output_path "./"

• For test set csv preparation: python ./misc_files/voc_to_csv_conversion.py --dataset_path "./PASCAL/Combined/" --dataset Test --output_path "./"

• Note: Here "Combined" directory contains two folders named "Train" and "Test" for voc_train and voc_test set.

3. Prepare tfrecords files from csv file

• For train set tfrecords: python ./misc_files/csv_to_tfrecords.py --output_path "./data/train.tfrecords" --csv_path "./data/ann_files/train_ann.csv" --num_splits 30

• For test set tfrecords: python ./misc_files/csv_to_tfrecords.py --output_path "./data/test.tfrecords" --csv_path "./data/ann_files/test_ann.csv" --num_splits 1

4. Prepare kmeans anchors for the dataset. python "./misc_files/get_kmeans_anchors_new.py" --train_csv_path "./data/ann_files/train_ann.csv" --output_anchor_path "./misc_files/anchors/kmeans_anchors_voc.txt"

• This will generate anchors for training set at the location specified by output_anchor_path.

5. [OPTIONAL] Convert darknet53 weights into tensorflow weights.

• Download weights from : https://pjreddie.com/media/files/darknet53.conv.74 and run below command. python "./pretrained_ckpt/convert_darknet53_weights.py" --darknet53_weights_path "./pretrained_ckpt/DarkNet53/darknet53_448.weights"

6. Modify config.py and run_model.py file as per the the training requirement and run below command. python run_model.py

Run Demo:

• For single image files modify and run test_on_image.py
• For video files modify and run test_on_video.py

File structre:

• models
  • Contains YoloV3 and TinyYoloV3 model definitions and loss function implementation.
• misc_files
  • Contains python files to prepare csv files and kmean anchors.
• train.py
  • Contains training scripts.
• inference_helper.py
  • Contains helper function to be used during the inference.
• config.py
  • Contains main important configuration for training. Before starting training you should be modifying this file as per your training requirement.
• dataset_helper.py
  • Contains data augmentation and dataset pipeline to get the augmented batch.
• run_classifier.py
  • To start training of Yolo model. Training will be done as per the config.py settings.

VOC 2007+12 Results:

• mAP @ 0.001 on VOC 07+12 train set AP: 73.70% (aeroplane) AP: 65.51% (bicycle) AP: 58.45% (bird) AP: 52.13% (boat) AP: 58.21% (bottle) AP: 74.69% (bus) AP: 67.97% (car) AP: 77.13% (cat) AP: 54.49% (chair) AP: 44.70% (cow) AP: 57.67% (diningtable) AP: 66.41% (dog) AP: 49.59% (horse) AP: 69.33% (motorbike) AP: 77.18% (person) AP: 46.18% (pottedplant) AP: 57.58% (sheep) AP: 55.59% (sofa) AP: 75.51% (train) AP: 59.58% (tvmonitor) mAP: 62.08%

• mAP @ 0.001 on VOC 07+12 test set AP: 62.84% (aeroplane) AP: 38.15% (bicycle) AP: 38.70% (bird) AP: 33.36% (boat) AP: 27.92% (bottle) AP: 64.53% (bus) AP: 61.82% (car) AP: 66.07% (cat) AP: 28.89% (chair) AP: 45.04% (cow) AP: 40.33% (diningtable) AP: 50.67% (dog) AP: 27.39% (horse) AP: 44.26% (motorbike) AP: 71.79% (person) AP: 24.77% (pottedplant) AP: 38.08% (sheep) AP: 40.60% (sofa) AP: 64.68% (train) AP: 36.29% (tvmonitor) mAP: 45.31%

Notes on obtained results:

• I have rigorously trained, tested and corrected this implementation. Although after so much of hardwork, I am not being able to completely reproduce the results of the paper. I have got 45.31 % mAP on VOC 07+12 test set and 62.08 % mAP on VOC 07+12 train set. The mAP on test set should be at least more than 70 %.
• So, if you are planning to use this implementation for some production work, please refer some other repos. This is not production ready implementation.

References:

• This implementation is highly inspired from Wizyoung's YoloV3 implementation. https://github.com/wizyoung/YOLOv3_TensorFlow
• Image augmentation using imgaug module https://github.com/aleju/imgaug
• mAP computation using below repo https://github.com/rafaelpadilla/Object-Detection-Metrics

Todos

• Multi-Scale Training
• Loss function implementation
• YoloV3 and TinyYoloV3 implementation
• Input pipeline using tf.data API
• Image augmentation using imgaug module
• Anchor calculation using kMeans Clustering