camera positioning and angle of view to get the suitable frame dimensions
we can simulate the view of camera on cad drawing to select the best position in order to get the view of a cylinder
we also can estimate the horizontal distance between the camera and the cylinder given the FOV and focal length
A frame from a camera should include the full view of a cylider and the cap of the cylinder
leakage of Data
we can select the suitable open source dataset
we can collect images from google then annotate it on roboflow website
we can select videos from youtube , split frames from it then start annotating it on roboflow
system work flow
we can put a marker on each slot for example (colours or characters) , rotate slots and apply a model on a stream video during slots rotatoin
pros
In this case we will increase our confidence in model prediction as we will use a voting technique (model will be applied on number of frames instead of one frame then it will return the most predicted class)
cons
In this case we will use 2 models (one for predicting the slot marker,and another one for predicting the cylinder type if existed) so, it will duplicate the inference time on each frame which is a constrain for us
we will face another problems in data collection for the model that is responsible for detecting the markers in slots
As the marker is fexed on slots , the existing of a cylinder may cause the disappearance of a marker (have to be checked on CAD)
Given a distance between slots from CAD , and RPM of motor we can calculate the working time of motor to put the slot in front of the camera , then call API to validate the presence of a camera and predict its type
pros
In this case we will apply the model on frame which will cause decreasing in inference time
cons
As we will apply a model on one frame , we cannot validate how the model works well
we can over come this issue by taking more than one frame (3 frames for example) and use voting technique
wa can put a QR code on each slot instead of markers
pros
easy in implementation
cons
QR code should be perpendicular on the camera position to be detected
Steps
Data Collection
Select the best model
Train & Validation
Test model
Deploy model on API as a "POC"
Continue the cycle till publishing the model into production to collect a real time data from site
Final Solution
Cylinders classifier
DONE
collect data for steel&fiber cylinders from Google, Facebook, Twitter, Snapchat, Roboflow and Youtube
Data collection
Data cleaning
Data annotating
select the best model
Yolov5 model for object detection problem
INCEPTIONV3 model for classification problem
Train&validate model
Slots
UNDER PROCESS
Generate characters data using word
Data generation
Data annotating
select the best model
Yolov5 model for object detection problem
VGG16 model for classification problem
Train, Validate model
Test model
cylinder caps classifier
UNDER PROCESS
collect data for cylinder caps from Google, Facebook, Twitter, Snapchat, Roboflow and Youtube but this feature will mainly depend on the collected data from real-time in the machine
Data collection
Data cleaning
Data annotating
select the best model
Yolov5 model for object detection problem
RESNET106 model for classification problem
Train&validate model
Deployment
UNDER PROCESS
Use Flask for API deployment
constrains
camera positioning and angle of view to get the suitable frame dimensions
what is the best type of camera
Leakage of Dataset
we have 6 doors..so, how will all systems work together?
time the the user should wait until the system response