Hi, we are working on recreating coral reef habitats from 20-min videos of the reef substrate taken using a fixed camera rig. The videos were audio synchronized and frames extracted at 3fps (native 24fps). This results in around 3900 images per camera (7800 images total) to be used for reconstruction. We use the sequential_matcher with the vocab tree to do feature matching and extracting. Then, to reduce processing time, we are using a sparse chunking method that implements image lists to split the images into more manageable chunks (9 total, 500 images each with 100 image overlap per camera) with image overlap among chunks that allows for sequential model_merging after sparse reconstruction (note: these are writing to the same .db). We do a bundle adjustment after each merge. We typically have great image registration at this stage (>90% in most cases) and the sparse reconstructions are sometimes good but for others we are encountering multiplanar reconstructions. And even in the better reconstructions, we are seeing slight duplication of structures, particularly near the edges, later in the dense reconstruction and dense meshing stages. We are wondering if there is a way to reduce these reprojection errors and the duplication of structures later in the pipeline and whether the model_merger step is the culprit. We tried GLOMAP reconstructions and were still running into the same multiplanar issues. We have also started a test model where there is an iterative point_triangulation and bundle_adjuster step following the final model_merge (noted in other threads as a possible fix), but this has yet to run through. Any additional advice on improving these models would be great!
Hi, we are working on recreating coral reef habitats from 20-min videos of the reef substrate taken using a fixed camera rig. The videos were audio synchronized and frames extracted at 3fps (native 24fps). This results in around 3900 images per camera (7800 images total) to be used for reconstruction. We use the sequential_matcher with the vocab tree to do feature matching and extracting. Then, to reduce processing time, we are using a sparse chunking method that implements image lists to split the images into more manageable chunks (9 total, 500 images each with 100 image overlap per camera) with image overlap among chunks that allows for sequential model_merging after sparse reconstruction (note: these are writing to the same .db). We do a bundle adjustment after each merge. We typically have great image registration at this stage (>90% in most cases) and the sparse reconstructions are sometimes good but for others we are encountering multiplanar reconstructions. And even in the better reconstructions, we are seeing slight duplication of structures, particularly near the edges, later in the dense reconstruction and dense meshing stages. We are wondering if there is a way to reduce these reprojection errors and the duplication of structures later in the pipeline and whether the model_merger step is the culprit. We tried GLOMAP reconstructions and were still running into the same multiplanar issues. We have also started a test model where there is an iterative point_triangulation and bundle_adjuster step following the final model_merge (noted in other threads as a possible fix), but this has yet to run through. Any additional advice on improving these models would be great!