ShishirPatil
commented
4 years ago Do you guys have the raw data used to train the model data.h on the github repository?
The GDPR compliant data set has been released, and can be accessed here - https://www.microsoft.com/en-us/research/uploads/prod/2018/05/dataTR_v1.tar.gz
What is the point of setting the toggle in the training pipeline to 0? Can the parts where the toggle is 0 not be discarded?
Think of toggle as a coarse label that can then be used to finely label the data. The coarse label is collected during data-collection, and the fine labeling is done post-data-collection using the scripts for labeling. For example, the toggle can be High for 5 seconds, and the gesture is a 2 second subset of the 5 seconds.
positive/negative examples used to train the model
The number of gestures of twirl being higher is an artifact of the process. There is nothing significant about this. Note that 353 includes chirality (right twirl and left twirl).
increasing the train set size by ten-fold
This is the idea of perturbation. Toy example: Consider a gesture that lasts 4 time stamps : [1,2,3,4,5,6,7,8,9,10] Here, {4,5,6,7} represent the gesture with {1,2,3,8,9,10} being the silence. Now I can generate multiple timestamp windows centered around the gesture {4,5,6,7}- a) {2,3,4,5,6,7,8}, b) {3,4,5,6,7,8,9}. The gestures only need to be approximately centered -there is some space to wriggle around.
users just walked around with the cane in hand and 8 minutes of that data was collected
Just use simple sliding windows to generate 2 second clips from the 8 minutes of data. Since, all the data is negative (no-gesture), just label all the 2 second windows as noise.
partial gestures. Should those examples also be labeled as noise?
Yes. Labeling partial gestures as noise helps increase precision.
For different data sizes, did the performance of the model not good enough until the last set of false positives were added?
The performance of model was "good" (high recall) even before the last set of false positives were added. The last set of false positives helped make the model more "robust" (lower false-positives) in our real-world deployment.
Hope this helps! Let me know if you have any further questions.
mousetrapbot01
Hello, I am trying to train new gestures for the GesturePod on the particle photon, a arduino-like mcu. I have some questions regarding training.
Do you guys have the raw data used to train the model data.h on the github repository?
What is the point of setting the toggle in the training pipeline to 0? Can the parts where the toggle is 0 not be discarded?
Also, the positive/negative examples used to train the model confuses me. If my understanding is correct, the positive examples were trained in a way where every gesture is performed within a 2 second slide, and then the user stops for a bit, in which case the label is 0, silence. Then the user performs it again. Why was the number of samples collected for twirl much higher 353 compared to the other gestures? Additionally, what did you mean by increasing the train set size by ten-fold by adding additional examples where the region is shifted by 25ms?
Also, for the negative examples, from my understanding, users just walked around with the cane in hand and 8 minutes of that data was collected. However, I don't understand where its says you guys clipped it to generate 2-second segments of negative examples. How do you label those negative examples? Should the negative examples be considered as noise for the 2 second segment and silence for the rest? Another confusion was about the partial gestures. Should those examples also be labeled as noise?
For different data sizes, did the performance of the model not good enough until the last set of false positives were added?
Thank you.