Page - 135 - in Joint Austrian Computer Vision and Robotics Workshop 2020

Objects Deformability Precision lemon juice bottle rigid 89.61 red tomatocan rigid 84.52 red teabox rigid 87.65 yellowtea box rigid 87.76 headphones slightlydef. 57.32 scissors slightlydef 54.15 humanhand model slightlydef. 37.16 powercable highlydef. 34.66 chain highlydef. 30.24 earphones highlydef. 14.86 Table1.Objectdetectionperformance, def -deformable andmanuallyannotated. Wethenran theYOLOnet- work trained with the synthetic data and calculated precision for each object, taking as ground truth the hand-annotated data. An Intersection Over Union (IOU)of50%wasconsideredasuccessfuldetection. As shown in previous work rigid objects like a can, a teaboxor lemon juicebottlehaveaverygood chanceatgettingdetectedwith theprecisionbeingat close to 90%. These objects do not change greatly inappearancewhenplaced indifferentpositionsand it is therefore easy for the network to learn their ap- pearance. We purposely choose that some of the ob- jectshavesimilarcolor, so that,due to lackaofgreat number of objects used for evaluation, the detection performance may not be attributed to simple color searching. Slightly deformable objects that we used were scissors, headphones and a human hand model. We see that in thecaseofslightlydeformableobjects the detection performance drops significantly with it be- ing around 55% for the scissors and the headphones. The chances of detecting the human hand model are even lower,being37.16%. The last threeobjects thatweevaluateareachain, apowercableandapairofearphones. Theseobjects are considered to be highly deformable. Again there is a clear drop in detection performance with the precisionofearphonesdetectionbeingonly14.86%. Chances that a powercableorapieceofachainwill bedetected are abit over 30%. All of the objects used for evaluation can be seen inFigure 1. Detectionofobjectsused forevaluation using YOLO trained on COCO dataset was unsuc- cessful for all of the objects except for the scissors with the detection rate of 62.35%, similar to our re- sult. The chain detection success cases can be seen onFigure 5,whereas thechain failurecasesarepre- sented inFigure 6. We can see that the in the cases where chain de- tection is successful a mask of chain taking a similar structure can be found in the bottom row of Figure 5. In the cases where the chain detection fails there are no masks available that resemble the given chain structure. We then pose the chain detection problem as sin- gle linkdetectionproblemandtry todetect thestruc- ture of the chain by detecting each individual link in the chain. In order to do so, we use our proposed method to segment the link in many different orien- tations and synthesize the training images. We then connect the individual links into a chain and test de- tection of individual links while the chain is taking different configurations. The results of a single link detectioncanbeseenon the toprowof theFigure 7. Figure7.Examples of linkdetection We also record 100 images of a chain taking dif- ferent shapes and manually annotate each of the links on the chain in all of the images and train the YOLOv3 network with those annotated images. The results of a single link detection with the manually annotated linkscanbeseenon thebottomrowof the Figure 7. We took chain as an example of a highly de- formable object that is made out of simple rigid ele- ments. Theseresultsshowthat thedetectionofande- formable object is possible by detecting its elemen- taryparts. Successfulandunsuccessfulcasesofobjectdetec- tion are presented in the Figure 4. As shown, on the examples of the power cable, the scissors and the headphones, detection is successful in some of the configurations. If the configuration is slightly changed the detection fails. This is due to the big variability in the appearance of these objects which is caused by their deformability. Potentially, mod- ellingofdeformableobjectssuchasapowercableor achaincouldbeused togeneratebigamountsofdif- 135