- Start a Virtual environment, you can run
python -m venv <path>. It is suggested that you use the name following virtual* pattern. For example, to create a venv in current directory, runpython -m venv . - If running this on Windows, you can launch the virtual environment using
.\Scripts\activate. - Install all dependencies using
pip install -r requirements.txt. We are using python 3.8 for this project. - If you run into any "Module not found" erros while running the code below, consider using
.\Scripts\pythonalias instead of simplypython. This is due some Windows system path incompatibilty. - [REQUIRED] Download the given dataset: https://github.com/FENGGENYU/PartNet_symh and let it reside at
dataset/Chair. Make sure that it's called exactly "Chair" and hasops/ syms/etc. folders within. - [REQUIRED] Download current ML model to \ML\ourML\checkpoint\tripleView: https://drive.google.com/file/d/1txSLUsvMYqhLJafw8QQPWzpFRICXC0zy/view?usp=sharing
- [OPTIONAL] Download precalculated joints folder to \grass-master\Chair\models\joints: https://drive.google.com/file/d/1bAlWhkQSQVjb75XzDEK7PfmdO-j73exm/view?usp=sharing
- Download final training data at https://drive.google.com/file/d/19w7n4ycMLvyRGky7xkI2eHFEkNHp4H6L/view?usp=sharing (not needed for execution)
- To test on setA, setB and setC, make sure to run the app like this:
python3 MixAndMatchApp.py -setB -g 5 -p 50. This will generate 50 chairs from set B and pick 5 best. - If you want to test it on a random collection, make sure to run the app like this:
python3 MixAndMatchApp.py -r 10 -g 5 -p 50. This will generate 50 chairs from random 10 and pick 5 best. - Not that the program exports generated in
.objformat inexport_objs/folder according to their index. 5 models would have indices 0, 1, 2, 3, 4 and they would correspond to0.obj,1.obj,2.obj,3.obj,4.obj.
- Run in the format of
python3 DatasetViewerApp.py n1 n2 n3...to view/operate specified dataset meshes. E.g.python3 DatasetViewerApp.py 2164 2165 2166 3452 5000 - Run in the format of
python3 DatasetViewerApp.py -r Xto view/operate random X meshes from the dataset. E.g.python3 DatasetViewerApp.py -r 10will pick 10 random meshes and put them up. - Run in the format of
python3 DatasetViewerApp.py -lto run the latest input. - Run in the format of
python3 DatasetViewerApp.py -setAto run one of the given data sets. - Press "a" and "d" to switch to prev/next models. Press "w" and "s" to look at particular parts of the model.
- Press "g" to generate a new chair. It's added in the end of all current models.
- Press "x" to take test depth screenshots of the chair present in the view. Saved to
screenshots/ - Press "e" to evaluate the chair in the view. The score will appear in the caption on the top.
- Press "y" to save chair as "positive". Requires presence of
dataset/imageData/chairs-data/positive - Press "n" to save chair as "negative". Requires presence of
dataset/imageData/chairs-data/negative
- By default the app will save the split objs to disk in a new folder, this was done for easier loading but can cause hard storage usage to go to 2x the base set when ran on every chair
- The first time a chair is used it trys to find connecting vertices based on distance from other meshes, these are then saved to disk, this slows down generation a bit. uploaded to google drive to avoid generation if wanted (generated with current distance threshold of 0.025 in getJoints)
- If MixAndMatch is ran with a large amount of generated chairs(>100 for setC) compared to the input size there are some outliers that are bad but scored as good that may show up frequently near the top -because of the above point the models are taken by generating 100 chairs from each set (aborting early if no new combinations are chosen for a number of selections)
- To re-generate the
dataset/compiledfolder, ensure that there isdataset/Chairfolder with all the dataset data within and rundataset/json-compile.pyfrom withindatasetfolder.
- Iavor: ML classificator
- Maheep: Mix and Matcher / Overall support
- Greg: Mix and Matcher
- Vlad: App maintenance / Overall support
- General Project Meeting - every Wednesday, 7:00 pm - 8:30 pm
- March 31st - Project Inception
- April 7th - Have dataset viewer ready to go
- April 14th - Crude and basic parts replacement, baseline sorting/scoring of the chairs
- April 21th - Refined parts replacement, scorer to sort the models
- April 26th, 12:30 pm - Project defense and submission (submission deadline is at 11:45 pm).
- Parse input meshes, initialize their joints and form collection of parts
- Pick random parts from the collection and weld them together using their joints. Do this for N chairs.
- Screenshot each chair and assign the score based on the depth screenshot. Sort by the score.
- Export and display.
- First milestone, this will require several other parts in the works
- Basic UX: "explore" the dataset, able to choose any provided mesh, able to look at the mesh and then look at each of the particular parts constituting the mesh
- Basic UI: one set of arrows to move between the meshes and another set of arrows to move between the parts in a particular mesh (or anything else simple and effective)
- Requires mesh-parts association which in turn requires dataset processing
- First working version of the project.
- Decent general functionality. Proper scoring and mesh welding.