Tensorflow realtime_object_detection on Jetson Xavier/TX2/TX1, PC

About this repository

forked from GustavZ/realtime_object_detection: https://github.com/GustavZ/realtime_object_detection
And focused on model split technique of ssd_mobilenet_v1.

Download model from here: tf1_detection_model_zoo

wget http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2018_01_28.tar.gz

and here: TensorFlow DeepLab Model Zoo

wget http://download.tensorflow.org/models/deeplabv3_mnv2_pascal_train_aug_2018_01_29.tar.gz

Support models

Model	model_type	split_shape
ssd_mobilenet_v1_coco_11_06_2017	nms_v0	1917
ssd_mobilenet_v1_coco_2017_11_17	nms_v1	1917
ssd_inception_v2_coco_2017_11_17	nms_v1	1917
ssd_mobilenet_v1_coco_2018_01_28	nms_v2	1917
ssdlite_mobilenet_v2_coco_2018_05_09	nms_v2	1917
ssd_inception_v2_coco_2018_01_28	nms_v2	1917
ssd_mobilenet_v1_quantized_300x300_coco14_sync_2018_07_03	nms_v2	1917
ssd_mobilenet_v1_0.75_depth_quantized_300x300_coco14_sync_2018_07_03	nms_v2	1917
ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03	nms_v2	51150
ssd_mobilenet_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03	nms_v2	51150
ssd_mobilenet_v1_ppn_shared_box_predictor_300x300_coco14_sync_2018_07_03	nms_v2	3000
faster_rcnn_inception_v2_coco_2018_01_28	faster_v2
faster_rcnn_resnet50_coco_2018_01_28	faster_v2
faster_rcnn_resnet101_coco_2018_01_28	faster_v2
faster_rcnn_inception_resnet_v2_atrous_coco_2018_01_28	faster_v2
mask_rcnn_inception_resnet_v2_atrous_coco_2018_01_28	mask_v1
mask_rcnn_inception_v2_coco_2018_01_28	mask_v1
mask_rcnn_resnet101_atrous_coco_2018_01_28	mask_v1
mask_rcnn_resnet50_atrous_coco_2018_01_28	mask_v1
deeplabv3_mnv2_pascal_train_aug_2018_01_29	deeplab_v3
deeplabv3_mnv2_pascal_trainval_2018_01_29	deeplab_v3
deeplabv3_pascal_train_aug_2018_01_04	deeplab_v3
deeplabv3_pascal_trainval_2018_01_04	deeplab_v3

• TensorRT -> model_type: 'trt_v1'
  Requirements: https://github.com/NVIDIA-Jetson/tf_trt_models
  

• Faster R-CNN: PC/Xavier only
  faster_rcnn_nas_coco_2018_01_28 occurred Out Of Memory on my PC.
  Other Faster R-CNN has not checked yet.
  

• Mask R-CNN: PC/Xavier only
  Removed split_model.
  Add worker_threads for parallel detection. A little bit fast, maybe.
  

• DeepLab V3: PC/Xavier only
  

See also:

• tensorflow/models#3270
• https://devblogs.nvidia.com/tensorrt-integration-speeds-tensorflow-inference/

Getting Started:

• login Jetson TX2. Desktop login or ssh remote login. ssh -C -Y ubuntu@xxx.xxx.xxx.xxx
• copy .config.yml to config.yml cp .config.yml config.yml
• edit config.yml for your environment. (Ex. camera_input: 0 # for PC)
• run python run_stream.py realtime object detection from webcam
• or run python run_video.py realtime object detection from movie file
• or run python run_image.py realtime object detection from image file
• wait a few minutes.
• Multi-Threading is better performance than Multi-Processing. Multi-Processing bottleneck is interprocess communication.
  

Requirements:

pip install --upgrade pyyaml

Also, OpenCV >= 3.1 and Tensorflow >= 1.4 (1.6 is good)

config.yml

Image

with run_image.py
Please create 'images' directory and put image files.(jpeg,jpg,png)
Subdirectories can also be used.

image_input: 'images'       # input image dir

Movie

with run_video.py

movie_input: 'input.mp4'    # mp4 or avi. Movie file.

Camera

with run_stream.py
This is OpenCV argument.

• USB Webcam on PC/Xavier

camera_input: 0

• USB Webcam on TX2

camera_input: 1

• Onboard camera on Xavier (with TX2 onboard camera)

camera_input: "nvarguscamerasrc ! video/x-raw(memory:NVMM), width=1280, height=720,format=NV12, framerate=120/1 ! nvvidconv ! video/x-raw,format=I420 ! videoflip method=rotate-180 ! appsink"

• Onboard camera on TX2

camera_input: "nvcamerasrc ! video/x-raw(memory:NVMM), width=(int)1280, height=(int)720,format=(string)I420, framerate=(fraction)30/1 ! nvvidconv flip-method=0 ! video/x-raw, format=(string)BGRx ! videoconvert ! video/x-raw, format=(string)BGR ! appsink"

Save to file

• Movie (run_stream.py or run_video.py)
  Save detection frame to movie file. (./output_movie/output_unixtime.avi)
  Requires a lot of disk space.
• Image (run_image.py)
  Save detection image to image file. (./output_image/PATH_TO_FILE/filename.jpg)
  Normally, this output image file is the same width x height and format as input images.
  But if run with MASK R-CNN, output file size is resized by width and height.

save_to_file: True

Without Visualization

I do not know why, but in TX2 force_gpu_compatible: True it will be faster.

• on TX2

force_gpu_compatible: True
visualize: False

• on PC

force_gpu_compatible: False
visualize: False

With Visualization

Visualization is heavy. Visualization FPS possible to limit.
Display FPS: Detection FPS.

• default is with Single-Processing and show every frames.

visualize: True
vis_worker: False
max_vis_fps: 0
vis_text: True

• Visualization FPS limit with Single-Processing

visualize: True
vis_worker: False
max_vis_fps: 30
vis_text: True

• Visualization FPS limit with Multi-Processing
  This is good to use with save_to_file: True.

visualize: True
vis_worker: True
max_vis_fps: 30
vis_text: True

• Model type

model_type: 'nms_v2'

The difference between 'nms_v1' and 'nms_v2' is BatchMultiClassNonMaxSuppression inputs.
model_type: trt_v1 is somewhat special. See config.yml.

# ssd_mobilenet_v1_coco_2018_01_28
model_type: 'nms_v2'
model_path: 'models/ssd_mobilenet_v1_coco_2018_01_28/frozen_inference_graph.pb'
label_path: 'models/labels/mscoco_label_map.pbtxt'
num_classes: 90

• Splite shape
  split_shape: 1917
  ExpandDims_1's shape. Ex:
  

learned size	split_shape
300x300	1917
400x400	3309
500x500	5118
600x600	7326

See also: Learn Split Model

• TensorRT split/non-split both support. Need Tensorflow with TensorRT support. (r1.9 has bug. I use r1.8/r1.10.1 for pc)

model_type: 'trt_v1'
precision_model: 'FP32'     # 'FP32', 'FP16', 'INT8'
model: 'ssd_inception_v2_coco_2018_01_28'
label_path: 'models/labels/mscoco_label_map.pbtxt'
num_classes: 90

Console Log

FPS:25.8  Frames:130 Seconds: 5.04248   | 1FRAME total: 0.11910   cap: 0.00013   gpu: 0.03837   cpu: 0.02768   lost: 0.05293   send: 0.03834   | VFPS:25.4  VFrames:128 VDrops: 1 

FPS: detection fps. average fps of fps_interval (5sec).
Frames: detection frames in fps_interval.
Seconds: fps_interval running time.

---

1FRAME
total: 1 frame's processing time. 0.1 means delay and 10 fps if it is single-threading(split_model: False). In multi-threading(split_model: True), this value means delay.
cap: time of capture camera image and transform for model input.
gpu: sess.run() time of gpu part.
cpu: sess.run() time of cpu part.
lost: time of overhead, something sleep etc.
send: time of multi-processing queue, block and pipe time.

---

VFPS: visualization fps.
VFrames: visualization frames in fps_interval.
VDrops: When multi-processing visualization is bottleneck, drops.

Updates:

• Support Xavier onboard camera. (with TX2 onboard camera)

• Add parallel detection for Mask R-CNN.

• Remove split from Mask R-CNN.

• Support DeepLab V3 models. model_type: deeplab_v3

• Add image input.

• Rename config.yml parameter name from save_to_movie to save_to_file.

• support Faster R-CNN models.

• Add max_frame: 0 for no exit with visualize: False.

• support ssd_mobilenet_v1 11 Jun, 2017 model.

• Add from movie.

• Add save_to_movie.

• BETA: Support MASK R-CNN models.

• Always split GPU/CPU device.

• Support SSD 2018_07_03 models.

• Support TensorRT Optimization. : Need TensorRT, Tensorflow with TensorRT.

• Support ssd_mobilenet_v2, ssdlite_mobilenet_v2 and ssd_inception_v2_coco. : Download model from here: detection_model_zoo

• Add Multi-Processing visualization. : Detection and visualization are asynchronous.

• Drop unused files.

• Add force_gpu_compatible option. : ssd_mobilenet_v1_coco 34.5 FPS without vizualization 1280x720 on TX2.

• Multi-Processing version corresponds to python 3.6 and python 2.7.

• Launch speed up. : Improve startup time from 90sec to 78sec.

• Add time details. : To understand the processing time well.

• Separate split and non-split code. : Remove unused session from split code.

• Remove Session from load frozen graph. : Reduction of memory usage.

• Flexible sleep_interval. : Maybe speed up on high performance PC.

• FPS separate to multi-processing. : Speed up.

• FPS streaming calculation. : Flat fps.

• FPS is average of fps_interval. : Flat fps.

• FPS updates every 0.2 sec. : Flat fps.

• solve: Multiple session cannot launch problem. tensorflow.python.framework.errors_impl.InternalError: Failed to create session.

My Setup:

• PC
  • CPU: i7-8700 3.20GHz 6-core 12-threads
  • GPU: NVIDIA GTX1060 6GB
  • MEMORY: 32GB
  • Ubuntu 16.04
    • docker-ce
    • nvidia-docker
      • nvidia/cuda
      • Pyton 2.7.12/OpenCV 3.4.1/Tensorflow 1.6.1
      • Pyton 3.6.5/OpenCV 3.4.1/Tensorflow 1.6.1
• Jetson Xavier
  • JetPack 4.0 Developer Preview
    • Python 2.7/OpenCV 3.3.1/Tensorflow 1.6.1
    • Python 2.7/OpenCV 3.3.1/Tensorflow 1.10.1 (slow)
  • JetPack 4.1.1 Developer Preview
    • Python 3.6.7/OpenCV 3.4.1/Tensorflow 1.10.1 (seems fast. I changed opencv build options.)
• Jetson TX2
  • JetPack 3.2/3.2.1
    • Python 3.6
    • OpenCV 3.4.1/Tensorflow 1.6.0
    • OpenCV 3.4.1/Tensorflow 1.6.1
    • OpenCV 3.4.1/Tensorflow 1.7.0 (slow)
    • OpenCV 3.4.1/Tensorflow 1.7.1 (slow)
    • OpenCV 3.4.1/Tensorflow 1.8.0 (slow)
  • JetPack 3.1
    • Python 3.6
    • OpenCV 3.3.1/Tensorflow 1.4.1
    • OpenCV 3.4.0/Tensorflow 1.5.0
    • OpenCV 3.4.1/Tensorflow 1.6.0
    • OpenCV 3.4.1/Tensorflow 1.6.1 (Main)
• Jetson TX1
  • SSD Storage
  • JetPack 3.2
    • Python 3.6
    • OpenCV 3.4.1/Tensorflow 1.6.0

NVPMODEL

Mode	Mode Name	Denver 2	Frequency	ARM A57	Frequency	GPU Frequency
0	Max-N	2	2.0 GHz	4	2.0 GHz	1.30 GHz
1	Max-Q	0		4	1.2 GHz	0.85 GHz
2	Max-P Core-All	2	1.4 GHz	4	1.4 GHz	1.12 GHz
3	Max-P ARM	0		4	2.0 GHz	1.12 GHz
4	Max-P Denver	2	2.0 GHz	0		1.12 GHz

Max-N

sudo nvpmodel -m 0
sudo ./jetson_clocks.sh

Max-P ARM(Default)

sudo nvpmodel -m 3
sudo ./jetson_clocks.sh

Show current mode

sudo nvpmodel -q --verbose

Current Max Performance of ssd_mobilenet_v1_coco_2018_01_28

FPS	Machine	Size	Split Model	Visualize	Mode	CPU	Watt	Ampere	Volt-Ampere	Model	classes
227	PC	160x120	True	False	-	27-33%	182W	1.82A	183VA	frozen_inference_graph.pb	90
223	PC	160x120	True	True, Worker 30 FPS Limit	-	28-36%	178W	1.77A	180VA	frozen_inference_graph.pb	90
213	PC	544x288	True	False	-	49-52%	178W	1.79A	180VA	frozen_inference_graph.pb	90
212	PC	160x120	True	True	-	30-34%	179W	1.82A	183VA	frozen_inference_graph.pb	90
207	PC	544x288	True	True, Worker 30 FPS Limit	-	48-53%	178W	1.76A	178VA	frozen_inference_graph.pb	90
190	PC	544x288	True	True	-	52-58%	176W	1.80A	177VA	frozen_inference_graph.pb	90
174	PC	1280x720	True	False	-	42-49%	172W	1.72A	174VA	frozen_inference_graph.pb	90
163	PC	1280x720	True	True, Worker 30 FPS Limit	-	47-53%	170W	1.69A	170VA	frozen_inference_graph.pb	90
153	PC	1280x720	True	True, Worker 60 FPS Limit	-	51-56%	174W	1.73A	173VA	frozen_inference_graph.pb	90
146	PC	1280x720	True	True, Worker No Limit (VFPS:67)	-	57-61%	173W	1.70A	174VA	frozen_inference_graph.pb	90
77	PC	1280x720	True	True	-	29-35%	142W	1.43A	144VA	frozen_inference_graph.pb	90
60	Xavier	160x120	True	False	Max-N	34-42%	31.7W	0.53A	54.5VA	frozen_inference_graph.pb	90
59	Xavier	544x288	True	False	Max-N	39-45%	31.8W	0.53A	54.4VA	frozen_inference_graph.pb	90
58	Xavier	1280x720	True	False	Max-N	38-48%	31.6W	0.53A	55.1VA	frozen_inference_graph.pb	90
54	Xavier	160x120	True	True	Max-N	39-44%	31.4W	0.52A	54.4VA	frozen_inference_graph.pb	90
52	Xavier	544x288	True	True	Max-N	39-50%	31.4W	0.55A	56.0VA	frozen_inference_graph.pb	90
48	Xavier	1280x720	True	True	Max-N	44-76%	32.5W	0.54A	55.6VA	frozen_inference_graph.pb	90
43	TX2	160x120	True	False	Max-N	65-76%	18.6W	0.28A	29.9VA	frozen_inference_graph.pb	90
40	TX2	544x288	True	False	Max-N	60-77%	18.0W	0.28A	29.8VA	frozen_inference_graph.pb	90
38	TX2	1280x720	True	False	Max-N	62-75%	17.7W	0.27A	29.2VA	frozen_inference_graph.pb	90
37	TX2	160x120	True	True	Max-N	5-68%	17.7W	0.27A	28.0VA	frozen_inference_graph.pb	90
37	TX2	160x120	True	False	Max-P ARM	80-86%	13.8W	0.22A	23.0VA	frozen_inference_graph.pb	90
37	TX2	160x120	True	True	Max-P ARM	77-80%	14.0W	0.22A	23.1VA	frozen_inference_graph.pb	90
35	TX2	544x288	True	True	Max-N	20-71%	17.0W	0.27A	27.7VA	frozen_inference_graph.pb	90
35	TX2	544x288	True	False	Max-P ARM	82-86%	13.6W	0.22A	22.8VA	frozen_inference_graph.pb	90
34	TX2	1280x720	True	False	Max-P ARM	82-87%	13.6W	0.21A	22.2VA	frozen_inference_graph.pb	90
32	TX2	544x288	True	True	Max-P ARM	79-85%	13.4W	0.21A	22.3VA	frozen_inference_graph.pb	90
31	TX2	1280x720	True	True	Max-N	46-75%	16.9W	0.26A	28.1VA	frozen_inference_graph.pb	90
27	TX1	160x120	True	False	-	71-80%	17.3W	0.27A	28.2VA	frozen_inference_graph.pb	90
26	TX2	1280x720	True	True	Max-P ARM	78-86%	12.6W	0.20A	21.2VA	frozen_inference_graph.pb	90
26	TX1	544x288	True	False	-	74-82%	17.2W	0.27A	29.0VA	frozen_inference_graph.pb	90
26	TX1	160x120	True	True	-	69-81%	17.1W	0.27A	28.7VA	frozen_inference_graph.pb	90
24	TX1	1280x720	True	False	-	73-80%	17.6W	0.27A	29.3VA6	frozen_inference_graph.pb	90
23	TX1	544x288	True	True	-	77-82%	16.7W	0.27A	28.2VA	frozen_inference_graph.pb	90
19	TX1	1280x720	True	True	-	78-86%	15.8W	0.26A	26.7VA	frozen_inference_graph.pb	90

on Xavier 544x288:

on PC 544x288:

on TX2 544x288:

Youtube

Robot Car and Realtime Object Detection

Object Detection vs Semantic Segmentation on TX2

Realtime Object Detection on TX2

Realtime Object Detection on TX1

Movie's FPS is little bit slow down. Because run ssd_movilenet_v1 with desktop capture.
Capture command:

gst-launch-1.0 -v ximagesrc use-damage=0 ! nvvidconv ! 'video/x-raw(memory:NVMM),alignment=(string)au,format=(string)I420,framerate=(fraction)25/1,pixel-aspect-ratio=(fraction)1/1' ! omxh264enc !  'video/x-h264,stream-format=(string)byte-stream' ! h264parse ! avimux ! filesink location=capture.avi

Training ssd_mobilenet with own data

https://github.com/naisy/train_ssd_mobilenet

Multi-Threading for Realtime Object Detection

Multi-Threading for Realtime Object Detection

Learn Split Model

Learn Split Model