- Time: 15:00 - 15:30
- Attendees: Bob Zhang (Supervisor), Huang Yanzhen, Mai Jiajun
- Revealed the excel, which lists different kinds of movement to be recorded as video(input). Including using postures and other extensions.
- Formulated the name of each video in shape of “YYYYMMDD_hhmm_name_labels_weight.mp4”, where “weight” can be decode as output y's . "weight" of each video is meant to be assigned manually.
- For the videos that have the property of chirality, we flip it vertically and consider it as a new video. We've argued about the correctness of the method of flipping video and finally convinced that it is okey for training.
- Discussed about transforming our classifier to the regressor.
- The new regressor is designed to estimate the "using" score of certain person (angles as input processed from skeleton landmarks).
- Output y's range from [0, 1] with type of float, the final score of one input will be close to 1.00 if the person is likely to using the phone, while 0.00 for not using on the contrary.
- Try to design a multi-layer structure of project.
- First layer : classify pedestrians into "using" and "not_using"(Binary Classification temporally)
- Next layer : use YOLOv5 or other tools to distinguish object phones around pedestrians labeled "using" in first layer.
- Third layer : use a more accurate method to make the judgement more persuasive.
- Share the experiment on workstation obtained from college CIS department.
- Follow the posture list from excel, record more video for training and testing.