Chunk alignment options - boat model

[ThomasVD]

Hi Agisoft Photoscan users!

I have a dilemma. I have generated a nice model of a historic sailing vessel, but I have generated it in 4 chunks: bow, stern, middle hull and sails. Now chunk alignment is finding the approximate positions, but it's not entirely perfect.

In order to combine the chunks accurately, and get the best possible result should I:
- export the 4 point clouds, align in meshlab, somehow import the aligned clouds back into Photoscan (will that still allow me to texturize them in Photoscan?)
- make a mesh of each individual chunk, tools -> import masks for each chunk, throw the masked photos into one chunk, and reprocess everything?
- not mask the individual chunks and just realign all pictures?

The reason I have processed in chunks is because there are several areas of the boat that look very alike (like stern and bow) but actually aren't. I'm afraid that if I process everything together, they will be misaligned (I had this previously on a boat that was very symmetrical, where Photoscan considered the starboard and port to be the same => it thought all pictures were taken from port rather than from both sides). Also, processing all together will take more computer time.

My thanks for any advice you may have!

Tom

[David Cockey]

I also create models of historic boats in multiple chunks. Maximum number of photos used for a project so far is around 800. My process is to align photos in chunks with as many photos as possible, and have photos shared between chunks. Then I align the chunks using the photo/camera method. I merge the chunks together and eliminate the photos. The result is a single chunk with all the photos and a sparse point cloud.

Chunks with dense point clouds and meshes are created starting the merged chunk with all the photos. For each new chunk I duplicate the merged chunk and then use the bounding box to set the size of the new chunk. The bounding box for each new chunk is set so there is a small amount of overlap between chunks. A dense point cloud and mesh is created for each chunk, and then the mesh is textured.

Attached is an image of a textured mesh with a set of lines overlaid. The mesh was created using the procedure described above. It was exported in .obj format and imported into Rhino where the lines were created. The mesh was not smoothed or edited other than trimming around the edges

[ThomasVD]

Hey David!

Thanks for the reply, your model looks amazing, great results!! I hadn't thought of aligning chunks with photo/camera align using "overlapping" pictures used in both chunks. Are the results in your experience better than with point chunk align? This sounds like a really great approach.

Just two small step that I'm not sure I understand, you said: "I merge the chunks together and eliminate the photos." Do you mean you eliminate the duplicates (ie the photos that are present in both chunks)?
After you have generated the separate dense point clouds and meshes, do you merge the meshes again before generating the final texture? Or do you essentially import several different texturized meshes into Rhino for creating the line plans etc?

Thanks again for your feedback, great to see how other people are using Photoscan for precisely the same purpose and getting awesome results!!

[David Cockey]

Yes, I delete duplicate photos.

I generally do not merge the dense point clouds and meshes in order to keep the sizes of the exported meshes from becoming too large (greater than 5 million faces).

[David Cockey]

The alignment of chunks which share photographs using the camera/photo method is generally as good or better than using the point method, and the camera/photo method takes only a few seconds compared to many minutes for the point method.