Object Detection on Raspberry Pi 4/3

Tested On

• RPi4 with USB camera
• RPi4 with Raspberry Pi Camera Module.
• RPi3 with USB camera
• RPi3 with Raspberry Pi Camera Module.

The purpose is to get the object detection and proof of concept working in the minimum time.

Ethos:

• We use pre-compiled binaries where possible from the Raspberry Pi repository.
• The python code contains the minimal needed to be functional.

Defaults

• Image size of 640x480
• ssdlite_mobilenet_v2_coco_2018_05_09

Rate

• RPi4 You can expect 2.3 FP/S using the defaults.
• RPi3 You can expect 1.2 FP/S using the defaults.

This uses pretrained models and can has the ability to change the model easy using the configuration file.

Scripts

To install run the following.

• To install Tensorflow 1

curl https://raw.githubusercontent.com/RattyDAVE/pi-object-detection/master/install.sh|/bin/sh

• To install Tensorflow 2 beta

curl https://raw.githubusercontent.com/RattyDAVE/pi-object-detection/master/install2.sh|/bin/sh

To start the object detetection run the following

curl https://raw.githubusercontent.com/RattyDAVE/pi-object-detection/master/run.sh|/bin/sh

To uninstall and wipe all traces run the following.

curl https://raw.githubusercontent.com/RattyDAVE/pi-object-detection/master/uninstall.sh|/bin/sh

Models

All models are taken from Tensorflow detection model zoo

The default is ssdlite_mobilenet_v2_coco_2018_05_09. You can change the models but uncommenting the line in the install.sh and the obj-config.ini

COCO-trained models from COCO dataset

90 Classes

Size	Model	Status rpi4	FPS rpi4
73MB	ssd_mobilenet_v1_coco_2018_01_28
44MB	ssd_mobilenet_v1_0.75_depth_300x300_coco14_sync_2018_07_03
81MB	ssd_mobilenet_v1_quantized_300x300_coco14_sync_2018_07_18
49MB	ssd_mobilenet_v1_0.75_depth_quantized_300x300_coco14_sync_2018_07_18
29MB	ssd_mobilenet_v1_ppn_shared_box_predictor_300x300_coco14_sync_2018_07_03
129MB	ssd_mobilenet_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03
349MB	ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03
179MB	ssd_mobilenet_v2_coco_2018_03_29
138MB	ssd_mobilenet_v2_quantized_300x300_coco_2019_01_03
48MB	ssdlite_mobilenet_v2_coco_2018_05_09	WORKS	2.8
265MB	ssd_inception_v2_coco_2018_01_28
142MB	faster_rcnn_inception_v2_coco_2018_01_28
363MB	faster_rcnn_resnet50_coco_2018_01_28
363MB	faster_rcnn_resnet50_lowproposals_coco_2018_01_28
622MB	rfcn_resnet101_coco_2018_01_28
565MB	faster_rcnn_resnet101_coco_2018_01_28
565MB	faster_rcnn_resnet101_lowproposals_coco_2018_01_28
641MB	faster_rcnn_inception_resnet_v2_atrous_coco_2018_01_28
641MB	faster_rcnn_inception_resnet_v2_atrous_lowproposals_coco_2018_01_28
1.09GB	faster_rcnn_nas_coco_2018_01_28
1.09GB	faster_rcnn_nas_lowproposals_coco_2018_01_28
693MB	mask_rcnn_inception_resnet_v2_atrous_coco_2018_01_28
169MB	mask_rcnn_inception_v2_coco_2018_01_28
631MB	mask_rcnn_resnet101_atrous_coco_2018_01_28
428MB	mask_rcnn_resnet50_atrous_coco_2018_01_28

Mobile models

Size	Model	Status rpi4	FPS rpi4
?	ssd_mobilenet_v3_large_coco
?	ssd_mobilenet_v3_small_coco

Pixel4 Edge TPU models

Size	Model	Status rpi4	FPS rpi4
?	ssd_mobilenet_edgetpu_coco

Kitti-trained models from Kitti dataset

2 Classes

Size	Model	Status rpi4	FPS rpi4
555MB	faster_rcnn_resnet101_kitti_2018_01_28	FAILED (bad alloc)

Open Images-trained models from Open Images dataset

601 Classes

Size	Model	Status rpi4	FPS 4
680MB	faster_rcnn_inception_resnet_v2_atrous_oid_2018_01_28
680MB	faster_rcnn_inception_resnet_v2_atrous_lowproposals_oid_2018_01_28
124MB	facessd_mobilenet_v2_quantized_320x320_open_image_v4
682MB	faster_rcnn_inception_resnet_v2_atrous_oid_v4_2018_12_12
151MB	ssd_mobilenet_v2_oid_v4_2018_12_12	Works	1.5
608MB	ssd_resnet101_v1_fpn_shared_box_predictor_oid_512x512_sync_2019_01_20

iNaturalist Species-trained models from iNaturalist Species Detection Dataset

2854 Classs

Size	Model	Status rpi4	FPS rpi4
868MB	faster_rcnn_resnet101_fgvc_2018_07_19
666MB	faster_rcnn_resnet50_fgvc_2018_07_19	FAILED (bad alloc)

AVA v2.1 trained models from AVA v2.1 dataset

AVA is a project that provides audiovisual annotations of video for improving our understanding of human activity.

90 Classes

Size	Model	Status rpi4	FPS rpi4
565MB	faster_rcnn_resnet101_ava_v2.1_2018_04_30	FAILED (bad alloc)

TypeError: int() argument must be a string, a bytes-like object or a number, not 'NoneType'