Seq_nms_YOLO

Introduction

This project combines YOLOv2(reference) and seq-nms(reference) to realise real time video detection.

Steps to execute

1. Download yolo.weights and tiny-yolo.weights by running

   wget https://pjreddie.com/media/files/yolo.weights
   wget https://pjreddie.com/media/files/yolov2-tiny-voc.weights -O tiny-yolo-voc.weights
   

2. While downloading the previous files create a Conda enviroment.

   conda create -y --name YOLOv2 python=2.7
   conda activate YOLOv2
   conda install -y -c conda-forge opencv
   conda install -y matplotlib Pillow scipy tensorflow
   conda install -y -c conda-forge tf_object_detection
   

3. Make sure that paths are correctly setted.

   export PATH=/usr/local/cuda-10.1/bin:$PATH
   

4. make the project.

   If you have a NVIDIA GPU you could set GPU=1. Also you might need to change the path of CUDA in lines 49 & 51. If don't have GPU just write GPU=0.

   If you want to use OpenCV set the flag OpenCV=1. If you want to use it but don't know how installed https://docs.opencv.org/3.3.1/d7/d9f/tutorial_linux_install.html. If you don't want to use it just set OpenCV=0.

5. Copy a video file to the video folder. (We added next videos as examples: v_ApplyEyeMakeup_g19_c03.avi, v_ApplyLipstick_g21_c01.avi, v_Archery_g07_c01.avi, v_BabyCrawling_g11_c01.avi, v_BalanceBeam_g18_c01.avi)

6. In the video folder, run python video2img.py -i input_file and then run python get_pkllist.py.

7. Return to root floder and run python yolo_seqnms.py to generate output images in video/output folder.

   From the label_map_util.py file in original repository we need to change tf.gfile.GFil by tf.io.gfile.GFile as the previous function is deprecated.

8. If you want to reconstruct a video from these output images, you can go to the video folder and run python img2video.py -i output.

   From the img2video.py file in original repository we changed cv2.cv.CV_FOURCC(*'mp4v') by cv2.VideoWriter_fourcc(*'mp4v') as the previous functions was not able. In the same file we commented cv2.destroyAllWindows() function as libgtk2.0-dev is not currently installed.

9. At last you will see output.mp4 file with detection results in video/ folder.

Use Yolov2 without Seq-NMS

In order to deactivate Seq-NMS post-processing add to the python yolo_seqnms.py execution --seq_nms 0 flag. If we want to deactivate the NMS delating non relevant detections add to the execution --nms 0 flag.

Reference

This project copies lots of code from darknet , Seq-NMS and models.

Videos come from UCF-101 dataset.