Ai python. -- 2
Budget: $8 – $15 USD
Inspired by the works referenced as MLMR, we seek to replicate the random forest classifier
workflow with different classes. Our requirements for this workflow are as follows:
• The input dataset will be .RCP / .E57
• A reformatting workflow will be proposed to transform the data into .TXT/.LAS or similar
• Point features should include X,Y,Z,R,G,B, normals & class as minimum
• The Random Forest classifier will have 3/4 layers/levels that will drill into greater details of
the dataset and more complex classes
• The training dataset will be only 10-15% of the full dataset
• The training dataset will be manually labelled
• The RF classifier will use the manually labelled dataset and perform the classification for the
remaining 85-90% of the dataset.
• A visualisation of the classified dataset will appear
• Each class will be saved separately
• Each class will save each instance separately.
We have found that point cloud files are huge in size and require a significant time to process hence
why the freelancer needs to take this into consideration. A small sample “room” file has been
provided with this brief, but this project should allow for input files of up to 500GB. A 200GB dataset
sample can be provided upon request. In general, we will like the freelance to design the code to
follow these user interaction touch points:
Firstly, the user will have to specify the location of the input point cloud file
• Then the user will conduct point cloud formatting & subsampling as input preparation
• Then the user will be prompted to select the training dataset (already classed) OR they will
have the option to label manually 10-15% of the original dataset (similar to this -
https://github.com/Hitachi-Automotive-And-Industry-Lab/semantic-segmentation-editor)
• Once the training data is added, the RF classifier performs the task
• The outcome is visualised in 3D
• The user moves onto a second level where they label more detailed features needed for
classification
• Once the training data is added, the RF classifier performs the task
• The outcome is visualised in 3D
• (these steps are repeated until the maximum level is reached)
• Classes are saved as individual files and then instances of each class are also saved
• Exported classes and segments are translated to .las/.rcp format
• A prompt appears confirming that the process has been undertaken correctly & accuracy /
F1 is confirmed
We don’t mind if this full solution is packaged into a software package or simply kept as a python
code. Nevertheless, a clear guide/manual will have to be provided with the code.
workflow with different classes. Our requirements for this workflow are as follows:
• The input dataset will be .RCP / .E57
• A reformatting workflow will be proposed to transform the data into .TXT/.LAS or similar
• Point features should include X,Y,Z,R,G,B, normals & class as minimum
• The Random Forest classifier will have 3/4 layers/levels that will drill into greater details of
the dataset and more complex classes
• The training dataset will be only 10-15% of the full dataset
• The training dataset will be manually labelled
• The RF classifier will use the manually labelled dataset and perform the classification for the
remaining 85-90% of the dataset.
• A visualisation of the classified dataset will appear
• Each class will be saved separately
• Each class will save each instance separately.
We have found that point cloud files are huge in size and require a significant time to process hence
why the freelancer needs to take this into consideration. A small sample “room” file has been
provided with this brief, but this project should allow for input files of up to 500GB. A 200GB dataset
sample can be provided upon request. In general, we will like the freelance to design the code to
follow these user interaction touch points:
Firstly, the user will have to specify the location of the input point cloud file
• Then the user will conduct point cloud formatting & subsampling as input preparation
• Then the user will be prompted to select the training dataset (already classed) OR they will
have the option to label manually 10-15% of the original dataset (similar to this -
https://github.com/Hitachi-Automotive-And-Industry-Lab/semantic-segmentation-editor)
• Once the training data is added, the RF classifier performs the task
• The outcome is visualised in 3D
• The user moves onto a second level where they label more detailed features needed for
classification
• Once the training data is added, the RF classifier performs the task
• The outcome is visualised in 3D
• (these steps are repeated until the maximum level is reached)
• Classes are saved as individual files and then instances of each class are also saved
• Exported classes and segments are translated to .las/.rcp format
• A prompt appears confirming that the process has been undertaken correctly & accuracy /
F1 is confirmed
We don’t mind if this full solution is packaged into a software package or simply kept as a python
code. Nevertheless, a clear guide/manual will have to be provided with the code.