This is the yolov3 person object detection framework which is based on the open-source keras yolov3 architecture.

The original project can be found in https://github.com/qqwweee/keras-yolo3.

###############################################################

#files description

##############################################################

--code

----keras-yolo3-master

------logs

--------000

----------trained_weights_final_2.h5 #final weight

----------ep004-loss36.290-val_loss34.747.h5 #middle weight

------model_data

--------coco_classes.txt #coco data label class

--------voc_classes.txt #voc data label class

--------yolo_anchors.txt #yolov3 anchor boxes size

--------yolo_weights.h5 #original yolov3 weight

------test

--------yolo3

----------init.py

----------model.py #yolov3 model

----------utils.py

--------1341846332.086502.png #object detection example of freiburg3

--------1341846337.150037.png #object detection example of freiburg3

--------halfsphere_bbox.txt #object detection result of freiburg3_walking_halfsphere data

--------show_pic.py #use the model to detect the freiburg3 example

--------xyz_bbox.txt #object detection result of freiburg3_walking_xyz data

--------yolo_copy.py #yolo_img_eval.py invoking function

--------yolo_img_eval.py #detect the object in freiburg3_walking_halfsphere and freiburg3_walking_xyz data

------yolo3

--------init.py

--------model.py #yolov3 model

--------utils.py

------coco_train.txt #coco 2014 train data labels

------yolo_anchors.txt #yolov3 anchor boxes size

------yolo.py #yolov3 detect object and function

------yolov3.cfg #yolov3 parameter

------2012_person_val.txt #voc 2012 validation data labels

------2012_person_train.txt #voc 2012 train data labels

------kmeans.py #k-means algorithm to get the suitable anchor boxes

------voc_annotation.py #get voc 2012 labels from original annotation format

------coco_voc.txt #voc 2012 and coco 2014 labels

------coco_val.txt #coco 2014 validation labels

------convert.py

------darknet53.cfg #darknet parameters

------train.py #train the yolov3 model

------coco_annotation.py #get coco 2014 labels from original annotation format

--data

----annotations

------person_keypoints_train2014.json #coco 2014 original labels(person train)

------person_keypoints_val2014.json #coco 2014 original labels(person validation)

----rgbd_dataset_freiburg3_walking_halfsphere

------rgb #freiburg3_walking_halfsphere rgb images

------depth #freiburg3_walking_halfsphere depth images

------accelerometer.txt

------depth.txt

------groundtruth.txt

------rgb.txt

----rgbd_dataset_freiburg3_walking_xyz

------rgb #freiburg3_walking_halfsphere rgb images

------depth #freiburg3_walking_halfsphere depth images

------accelerometer.txt

------depth.txt

------groundtruth.txt

------rgb.txt

----train2014 #coco 2014 images

----VOCdevkit

------VOC2012

--------Annotations #all voc xml labels

--------ImageSets

----------main

------------person_trainval.txt #voc person image IDs

------------person_train.txt #voc person train imgae IDs

------------person_val.txt #voc person validation image IDs

--------JPEGImages #voc images

--------SegmentationClass

--------SegmentationObject

################################################################

#code opration method

################################################################

1. run "voc_annotation.py" #get voc person labels, store them in "2012_person_val.txt" and "2012_person_train.txt". run "coco_annotation.py" #get coco person labels, store them in "coco_train.txt" and "coco_val.txt".

2. convert the "2012_person_val.txt" and the "coco_train.txt", store them in "coco_voc.txt".

3. run "k-means.py" to get the suitable anchor boxes size, which will use "coco_voc.txt".

4. run "train.py" to train the model. #"train.py" will use "trained_weights.h5", "yolo3","coco_voc.txt","voc_classes.txt","yolo_anchors.txt" and store the new weights

5. run "test/yolo_img_eval.py", use the trained model to detect the freiburg3 dataset and store the result in "halfsphere_bbox.txt" and "halfsphere_bbox.txt". run "test/show_pic.py" to show object detection example in freiburg3 dataset.

############################

ps: The data file does not have the image data, you can download the "VOCtrainval_11-May-2012.tar" from voc webpage, "annotations_trainval2014.zip" and "train2014.zip" in coco web page, "rgbd_dataset_freiburg3_walking_halfsphere.tgz" and "rgbd_dataset_freiburg3_walking_xyz.tgz" from freiburg3 web page. Then un compress them in the "data" folder.