Object detection for World Robot Summit

TBD

code repository
https://github.com/dongjuns/WRS_ObjectDetection

wrs_coco format
README.MD

pre-print for KROC or else
training phase plot
validation
mAP

code clean up and commit

0. Overview

Detecting the WRS objects for the robot arm to solve pick & place task

1. Dataset

https://github.com/RasmusHaugaard/wrs-data-collection
object classes: 12, 1920 x 1200 pixels, RGBA 4 channels, 250 images
strategy: train 200 images, validation 50 images

2. Object Detection

EfficientDet D0~D7

3. Get the bbox position of the rubber band

3-1. Remove the color threshold for just orange rubber band.
3-2. straight-forward solution: using blender simulation environment, set the camera paremeters and size, position and objects also.

4. Object detection using real time input image

5. Results

Best result: EfficientDet-D4, mAP 0.866

Model	obj1	obj2	obj3	obj4	obj5	obj6	obj7	obj8	obj9	obj10	obj11	obj12	All mAP	FPS
EfficientDet-D0	0	0.77	0.5	0.63	0.86	0.77	0.24	0.19	0.12	0.07	0.02	0.007	0.349	61
EfficientDet-D1	0.38	0.98	0.66	0.89	0.94	0.91	0.63	0.59	0.52	0.26	0.06	0.005	0.569	50
EfficientDet-D2	0.87	0.99	0.68	0.96	0.95	0.92	0.77	0.71	0.85	0.68	0.43	0.14	0.746	44
EfficientDet-D3	0.97	0.99	0.77	0.96	0.97	0.94	0.83	0.79	0.85	0.76	0.79	0.59	0.853	31
EfficientDet-D4	0.99	1	0.48	0.96	0.99	0.97	0.88	0.81	0.9	0.83	0.86	0.74	0.866	21
EfficientDet-D5	1	0.76	0.54	0.62	0.97	0.98	0.47	0.69	0.91	0.79	0.82	0.68	0.769	13
EfficientDet-D6	0.99	0.64	0.51	0.49	0.99	0.98	0.35	0.77	0.92	0.78	0.77	0.68	0.74	10
EfficientDet-D7	0.99	0.99	0.73	0.54	0.98	0.95	0.38	0.78	0.91	0.82	0.83	0.66	0.797	7