3D Reconstruction: From Pictures to Point Cloud

In this article, I will present you the different tools that we used to perform 3D reconstruction. In fact, if we take several pictures of an object, we are able to create a point cloud that represents this latter, in 3D. Here, we will not speak about Photogrammetry and Structure from Motion, because it will be the subject of the next article.

Here, we will first present Meshlab, the software used to vizualise and modify 3D object. Then, we will present you two tools that are used to reconstruct a model thanks to various pictures. I will also speak a little bit about Cloud Compare, another software that can be useful.

---

1) Meshlab

Meshlab is a software that is used to render, and modify 3D models. It is very useful, intuitive, and we will open it many times while doing the reconstruction. In the laboratory session, we learned basic operations on meshes. We downloaded models like a rabbit or a dragon, and tried to perform simple transformations. We were able to move, rotate, and scale them. We were also able to delete some points or vertices, which can be really important if we want to clear 3D reconstructed structures.

Where are the ears ? Deleted !

We can also compute the different normals (this is useful if we want to create meshes from a point cloud), and finally, we tried different simplifications (we delete points and recreate new meshes, in order to perform a glob reduction of the file's size).

Simplification of a Dragon

During this session, we also tried to use the PyMeshlab module in Python. It allows us to perform the same changes as before, but in Python. Indeed, we can do different changes at the same time on several images simultaneous.

In class, we were able to change the color of a 3D bunny:

We finally tried to combine the different filters available:

Here, we tried to change the color, the brightness, but also to simplify the model.

---

2) 3D Reconstruction

Now that we understood how to use the Meshlab software, we tried a more intersting part: 3D reconstruction.

Online Platform

The first tool was an online platform that is really efficient and simple to use. We just have to put around 20 images on a project, and the tool will automatically compute the point cloud. It took 20 minutes with this small quantity of images. Here is some of the photographies that I used:

At this end, I obtained this 3D model:

The result is quite convincing ! We can improve it by avoiding reflection problems, but here we can recognize the object. I also tried what we have seen in the first part to clean the result. At the end, we obtained this result:

Colmap

Now, we can try to perform a 3D reconstruction, but with the Colmap software. We can have great results too, but it is faster than the online one. In this one, I used a dataset with pictures of a building. Here, I was able to use 100 images (more details, a bigger object). Here are some examples:

At the end, we obtain this point cloud:

The result is great for a great building like this one. We can also try to clean a little bit, and we obtain that:

This is really efficient !

---

3) Cloud Compare

Cloud Compare is a software like Meshlab, but you have other tools in this one. For this work, I imported the model of a red dressed woman:

Now, we can try some actions, like some transformations (translations, rotations or scaling). We can compute the Interframe prediction (distance between the two positions), and see that visually:

We also tried, from a point cloud, to obtain a mesh. Here is the result:

Finally, we tested the SOR filter, which is capable of eliminating the points that are far away from the structure, and put them in the mesh. It is great for a cleaning purpose. I tried it with the building from Colmap, and I obtained that:

To conclude, we studied many tools to create 3D reconstruction, and to modify them. In the photogrammetry article, we will see a new code for this purpose.