Skip to content

Latest commit

 

History

History
149 lines (121 loc) · 5.99 KB

SSD_Tracking-Report.md

File metadata and controls

149 lines (121 loc) · 5.99 KB
Raw

This document contains video performance results for SSD_Tracker.

Test Environments:

  1. Local Machine

    1. Operating System: Ubuntu 16.04.4 LTS
    2. OpenCV version: 3.3.1
    3. GPU: GeForce GTX 645/PCIe/SSE2
    4. CPU: Intel® Core™ i7-4770 CPU @ 3.40GHz × 8
    5. CUDA version: 8.0
    6. cuDNN version: 6.0

  2. Virtual Machine

    1. Operating System: Ubuntu 16.04.4 LTS
    2. OpenCV version: 3.4.1
    3. GPU: NVIDIA M60 TESLA
    4. CPU: Intel(R) Xeon(R) CPU E5-2690 v3 @ 2.60GHz
    5. CUDA version: 8.0
    6. cuDNN version: 6.0

Test 1:

In this test, SSD detection and tracking is run on the same video with different resolution on virtual machine and local machine. The Total time is recoded for each video which includes

  1. ssd detection
  2. tracking
  3. drawing on boxes and tracking lines on the frames
  4. storing output video on disk.

Frame rate is calculated by dividing Total time by number of frames.

Video Link: https://www.youtube.com/watch?v=gBk6oT8YrJQ
Downloader: QDownloader.net
Video Length: 13 seconds (first 400 frames) Video Frame Rate: 30 fps
SSD Detection: models_VGGNet_VOC0712_SSD_300x300 video output fcc: fourcc('M', 'J', 'P', 'G')

Note: the video is downloaded without sound.

  1. Virtual Machine

    1. Resolution: 1920 x 1080
      Total time = 20.2323 seconds | Frame rate: 19.8198 fps
    2. Resolution: 1280 x 720
      Total time = 15.4806 seconds | Frame rate: 25.9034 fps
    3. Resolution: 854 x 480
      Total time = 13.2859 seconds | Frame rate: 30.1824 fps
    4. Resolution: 426 x 240
      Total time = 11.8224 seconds | Frame rate: 33.9188 fps

  2. Local Machine

    1. Resolution: 1920 x 1080
      Total time = 105.551 seconds | Frame rate: 3.79913 fps
    2. Resolution: 1280 x 720
      Total time = 102.063 seconds | Frame rate: 3.92895 fps
    3. Resolution: 854 x 480
      Total time = 98.7727 seconds | Frame rate: 4.05983 fps
    4. Resolution: 426 x 240
      Total time = 98.1054 seconds| Frame rate: 4.08744 fps

Conclusion:

Virtual machine performs better, as expected compare to local machine. The detection time per frame decreases with decrease in video resolution. There is a more drastic decrease in frame rate when only detection is performed.

Please also refer to the following test(s)

  1. Caffe test

Test 2:

In this test, SSD detection and tracking is run on the same video with different resolution on virtual machine. The Total time is recoded for each video which includes

  1. ssd detection
  2. tracking
  3. drawing on boxes and tracking lines on the frames
  4. storing output video on disk.

Frame rate is calculated by dividing Total time by number of frames.

Video Link: https://www.youtube.com/watch?v=gBk6oT8YrJQ
Downloader: QDownloader.net
Video Length: 13 seconds (first 400 frames) Video Frame Rate: 30 fps
SSD Detection: models_VGGNet_VOC0712_SSD_300x300 fcc video: cv::VideoWriter::fourcc('M', 'J', 'P', 'G')

Note: the video is downloaded without sound.

  1. Virtual Machine

    1. Resolution: 1920 x 1080
      Total time = 20.3877 seconds | Frame rate: 19.6688 fps
      Total Detection time = 11.5326 seconds
      Total Tracking time = 2.86798 seconds
      Total Drawing time = 0.149267 seconds
      Total Storing time = 4.10904 seconds

    2. Resolution: 1280 x 720
      Total time = 15.5311 seconds | Frame rate: 25.8192 fps
      Total Detection time = 11.4366 seconds
      Total Tracking time = 1.04571 seconds
      Total Drawing time = 0.115335 seconds
      Total Storing time = 2.13758 seconds

    3. Resolution: 854 x 480
      Total time = 13.4501 seconds | Frame rate: 29.8138 fps
      Total Detection time = 11.3153 seconds
      Total Tracking time = 0.438523 seconds
      Total Drawing time = 0.102125 seconds
      Total Storing time = 1.15259 seconds

    4. Resolution: 426 x 240
      Total time = 11.901 seconds | Frame rate: 33.6946 fps
      Total Detection time = 11.1076 seconds
      Total Tracking time = 0.114087 seconds
      Total Drawing time = 0.0736228 seconds
      Total Storing time = 0.431988 seconds

Conclusion:

Detection, tracking, drawing and storing time increases with increase in resolution. Storing time seems to increase the most. The output file is very big in size. The original 1920 x 1080 video is 85.9 MB and the 13 second output is 128.2MB. As the writing file is huge, it is taking a long time to write it to the disk.

Please also refer to the following test(s)

  1. Caffe test

Test 3:

In this test, Output video size and storing time is recorded for different fourcc value. The frame goes through drawing part where it is modified and then stored to the disk.\

fourcc value: - 4-character code of the codec which is used to compress the video.\

Video Link: https://www.youtube.com/watch?v=gBk6oT8YrJQ
Downloader: QDownloader.net
Video Length: 13 seconds (first 400 frames) Video Frame Rate: 30 fps
SSD Detection: models_VGGNet_VOC0712_SSD_300x300

Note: the video is downloaded without sound.

  1. Virtual Machine
    1. VideoWriter::fourcc('P', 'I', 'M', '1') for MPEG-1
      Total Storing time = 9.28308 seconds | Size: 34.7 MB
    2. VideoWriter::fourcc('M', 'J', 'P', 'G') for Motion JPEG Total Storing time = 4.22593 seconds | Size: 129.0 MB
    3. VideoWriter::fourcc('M', 'P', '4', '2') for MPEG-4 variation of Microsoft Total Storing time = 7.21563 seconds | Size: 41.6 MB

Conclusion:

Video output writing depends on the encoding used. Further test will be conducted once the input video format is known.

Please also refer to the following test(s)

  1. Caffe test