Fastest Framework for Object detection on Jetson TK1?

[Gemm]

Hi,

Has anyone bench marked the execution time and frames per second for object detection application with different deep learning frameworks?

Which one among the Deep Learning frameworks (Caffe, Torch, Darknet) might be the fastest for object detection and localization with bounding box prediction for real-time video on Jetson TK1?

Thanks,

[AastaLLL]

Hi BharatS,

Here are some data tested on tx1 for your reference.

• Faster-RCNN(Caffe): GitHub - rbgirshick/py-faster-rcnn: Faster R-CNN (Python implementation) -- see https://github.com/ShaoqingRen/faster_rcnn for the official MATLAB version
  Run tools/demo.py

Loaded network /home/ubuntu/JEP/topic_974063/py-faster-rcnn/data/faster_rcnn_models/VGG16_faster_rcnn_final.caffemodel
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/000456.jpg
Detection took 1.641s for 300 object proposals
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/000542.jpg
Detection took 1.467s for 161 object proposals
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/001150.jpg
Detection took 1.623s for 194 object proposals
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/001763.jpg
Detection took 1.657s for 196 object proposals
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Demo for data/demo/004545.jpg
Detection took 1.617s for 300 object proposals

• Detectnet+tensorRT: GitHub - dusty-nv/jetson-inference: Hello AI World guide to deploying deep-learning inference networks and deep vision primitives with TensorRT and NVIDIA Jetson.
  Run ./detectnet-console peds-007.png output-7.png
  <This time only take inference time, no image read/write included.>

detectnet-console:  take 0.021434 for inferencing
9 bounding boxes detected
bounding box 0   (728.078613, 223.725586)  (802.968750, 359.945068)  w=74.890137  h=136.219482
bounding box 1   (399.096680, 284.106445)  (471.357422, 502.163086)  w=72.260742  h=218.056641
bounding box 2   (621.679688, 303.618164)  (715.136719, 765.043945)  w=93.457031  h=461.425781
bounding box 3   (1198.300781, 376.259766)  (1310.361328, 655.290527)  w=112.060547  h=279.030762
bounding box 4   (778.901367, 522.136230)  (898.505859, 859.042969)  w=119.604492  h=336.906738
bounding box 5   (459.645996, 564.719238)  (530.610352, 746.718750)  w=70.964355  h=181.999512
bounding box 6   (929.677734, 798.859863)  (1065.615234, 1108.872070)  w=135.937500  h=310.012207
bounding box 7   (1240.004883, 865.964355)  (1381.816406, 1159.497070)  w=141.811523  h=293.532715
bounding box 8   (729.492188, 883.300781)  (871.054688, 1240.839844)  w=141.562500  h=357.539062
draw boxes  9  0   0.000000 200.000000 255.000000 100.000000
detectnet-console:  writing 1920x1080 image to 'output-7.png'
detectnet-console:  successfully wrote 1920x1080 image to 'output-7.png'

shutting down...

• YOLO: YOLO: Real-Time Object Detection
  Run ./darknet yolo test cfg/yolo-small.cfg yolo-small.weights
  

Loading weights from yolo-small.weights...Done!
Enter Image Path: peds-007.png
peds-007.png: Predicted in 21.923227 seconds.
person: 0.39
person: 0.20
person: 0.23
person: 0.29
Not compiled with OpenCV, saving to predictions.png instead
Not compiled with OpenCV, saving to resized.png instead