No speed improvement with multi-GPU when extracting and matching SIFT features

[Wayne Cochran]

I am on a machine we call the beast which has 8 Quadro P5000 which each have 16GB of RAM. The machine itself has 48 cores (with hyper threading) Intel Xeon Gold 6146 CPU @ 3.20GHz over 700GB RAM .  Running Linux/Ubunu 16.04.

I have performed a SfM sparse reconstruction of a scene with 155 6K images and it works beautifully and I don't care too much about speed here.

Later I want to register a single 2K image with this scene and this part I want to be fast. I have been doing this using the sequence of colmap commands using the feature_extractor, exhaustive_matcher, and image_registrator. To my joy I saw there was a multi-gpu option for the feature_extractor and feature_matcher and decided to greedily use all my GPU's, but alas there was no speed-up?  Below is my script. The file reg-image-list.txt references a single 2K image. This takes about 25 seconds whether I use the multi-gnu option or not -- why isn't there a speed up?

time colmap feature_extractor \

   --database_path $DB \

   --image_path reg-images \

   --image_list_path reg-image-list.txt \

   --ImageReader.camera_model SIMPLE_PINHOLE \

   --ImageReader.camera_params "$F, $CX, $CY" \

   --SiftExtraction.use_gpu 1 \

   --SiftExtraction.gpu_index=0,1,2,3,4,5,6,7 \

   --SiftExtraction.domain_size_pooling 1 \

   --SiftExtraction.estimate_affine_shape 1

time colmap exhaustive_matcher \

     --database_path $DB \

     --SiftMatching.use_gpu 1 \

     --SiftMatching.gpu_index=0,1,2,3,4,5,6,7

time colmap image_registrator \

    --database_path $DB \

    --input_path  $IN_MODEL \

    --output_path  $OUT_MODEL \

    --Mapper.ba_local_max_num_iterations 40 \

    --Mapper.ba_global_max_num_iterations 100 \

    --Mapper.ba_local_max_refinements 3 \

    --Mapper.ba_global_max_refinements 5

[Revisionarian]

The GPU version of feature_extractor does not support the "domain_size_pooling" or "estimate_affine_shape" options. So COLMAP falls back to the CPU version for feature extraction in this case.