Tables

*  The metrics «APK and PCK» are expressed in «%», whereas the «error» is expressed in «pixels».

 

  • The following table, shows the results obtained to decide the correct operation of the background removal using MSER. To do this, the 4D-DPM model was trained using a set of images from the extended CAD60 database to validate the use of the MSER
 Modelo  MSER training  MSER testing  Eval.  HEAD  SHOULD.  WRIST  HIP ANKLE AVG.
1

no

 

 no APK  100 100 89.64 100 100 97.92
PCK 100 100 92.42 100 100 98.48
error 4.22 3.66 7.63 5.96 4.43 5.18
2 no si APK 100 100 83.79 100 100 96.75
PCK  100 100 89.39 100 100 97.87
error  4.57 3.61 7.70 3.35 3.77 4.59
3 si no APK  100 100 82.40 100 100 96.49
PCK  100 100 87.37 100 100 97.47
error  3.03 4.49 9.65 3.38 3.09 4.72
4 si si APK 100 100 95.63 100 100 99.12
PCK 100 100 96.46 100 100 99.29
error 2.55 4.70 5.62 3.41 2.64 3.78

 

  • The original DPM model is trained with the database «PARSE», which only contains images in RGB, so that it can not be compared with the proposed 4D-DPM. To do this, we re-train the original DPM model with the «CAD60 dataset» and train the proposed 4D-DPM model. To do this, the same images are used to train both models and the same images are used to test the obtained models, but the training images and the test images are not the same. The resulting table is the next table. For a correct comparison of both methods in the 4D-DPM model, the particle filter described and the MSER method for the removal of the bottom have not been used. The table displays the 10 parts used in the proposed 4D-DPM model, comparing them to the same 10 parts of the original DPM model, although the original model contains 14 parts. These 10 parts are: the head, shoulders, hands, hips and ankle. The elbows and knees are not compared because they will be compared later when using the dual quaternions to infer in the number of points used. At these points the trunk position is added, but this point is not used in the tables.
MODEL EVAL. HEAD SHOULD. WRIST HIP ANKLE AVG.
DPM APK 47.42 66.69 22.95 45.98 47.10 46.02
PCK 62.00 70.50 39.00 60.00 57.50 57.80
error 17.35 14.10 35.89 7.06 19.57 18.79
DPM-t APK 73.02 73.53 35.26 66.33 42.38 57.50
PCK 78.50 78.50 44.50 70.50 49.50 64.30
error 15.21 12.30 31.02 6.64 16.31 16.29
4D-DPM APK 91.23 87.06 51.63 86.21 82.01 79.63
PCK 92.80 90.00 66.00 89.00 90.00 85.56
error 8.81 7.53 19.25 6.05 9.25 10.17

 

  • The first idea used to improve the results of the 4D-DPM model was to introduce the Kalman filter to verify that the results improved and to be able to subsequently study other options to track the points of interest within an image. The next table shows the results obtained.

Where «DPM-t» Represents the original DPM model trained and tested in the «CAD60» dataset. «4D-DPM» Represents the proposed model trained and tested in the database CAD60? In addition to the two solutions provided, with and without the Kalman filter. To obtain these results in both models, the same images have been used to perform the training and testing of the model, but the training and testing images are different from each other. Observing the next table, we return to the same conclusion as before, that with the 4D-DPM model, a higher accuracy is obtained using and without using the Kalmans filter, in addition to the accuracy of the 4D-DPM model using the Kalman filter provides 3.5% more accuracy compared to the same model without using the Kalman filter.

MODEL EVAL. HEAD SHOULD. WRIST HIP ANKLE AVG.
DPM-t APK 91.20 92.30 82.70 86.60 83.50 87.26
PCK 91.50 89.00 85.80 89.90 83.80 88.00
error 8.17 8.81 10.87 9.37 11.59 9.76
4D-DPM
without KF		 APK 94.20 95.10 88.30 89.70 90.30 91.52
PCK 93.80 92.50 88.90 90.30 91.00 91.30
error 6.48 6.02 8.73 8.01 7.66 7.38
4D-DPM
with KF		 APK 97.50 98.30 92.20 94.70 94.00 95.34
PCK 96.40 95.20 93.70 96.50 94.20 95.20
error 5.82 5.71 7.43 6.37 6.61 6.38

 

  • After observing the improvement when introducing the Kalman filter, it is proposed to introduce a different filter, in this case the particle filter, since these types of filters get better responses in nonlinear systems as in our case. Next table shows the results of the 4D-DPM model with and without the use of the Particle Filter
MODEL EVAL. HEAD SHOULD. WRIST HIP ANKLE AVG.
4D-DPM
without PF		 APK 92.60 93.20 87.10 88.20 88.50 89.92
PCK 93.10 91.70 86.70 89.40 90.40 90.56
error 6.95 7.15 9.58 8.43 7.89 8.00
4D-DPM
with PF		 APK 94.10 95.40 90.40 91.20 91.60 92.54
PCK 96.40 92.90 91.10 92.00 92.10 92.10
error 6.33 6.97 7.48 7.25 7.35 7.07

 

  • After verifying that both the Kalman Filter and the Particle Filter improve the results of the original DPM model, by making the 4D-DPM model obtain better precision in the results, a study is carried out to verify which filter performs an improvement more substantial in the results. To do this, the following table is displayed:
MODEL EVAL. HEAD SHOULD. WRIST HIP ANKLE AVG.
4D-DPM
with KF		 APK 89.60 88.70 84.20 87.50 85.90 97.18
PCK 90.10 88.90 85.20 88.00 87.10 87.86
error 8.16 8.67 9.94 9.25 8.97 8.99
4D-DPM
with PF		 APK 92.30 91.10 87.60 89.50 88.20 89.74
PCK 92.80 90.30 88.10 90.70 89.90 92.10
error 7.15 8.02 8.36 8.44 8.69 8.13

 

  • Next, the «4D-DPM» model is compared, subtracting the background using MSER and using the particle filter to increase sensor accuracy with the original «DPM-t» model, where only RGB images are used and with the algorithm from «Kinect», where only depth images are used.
MODEL EVAL. HEAD SHOULD. WRIST HIP ANKLE AVG.
DPM-t APK 47.30 66.70 22.40 45.50 47.10 45.80
PCK 62.50 70.40 39.00 60.50 57.90 58.06
error 15.53 12.23 22.34 16.29 18.50 16.97
Kinect APK 68.30 90.70 76.40 9.50 77.10 64.40
PCK 79.50 94.40 85.00 23.50 85.90 73.66
error 13.17 6.85 9.64 18.42 11.28 11.87
4D-DPM APK 75.40 93.00 83.70 85.50 84.20 84.36
PCK 84.10 96.30 90.20 88.90 89.90 89.88
error 10.59 5.98 8.13 9.82 9.08 8.72