Skip to content

Latest commit

 

History

History
68 lines (48 loc) · 2.98 KB

image_capture.md

File metadata and controls

68 lines (48 loc) · 2.98 KB
Performance Benchmarks

Image capture

These benchmarks measure the FPS of image capture and observation data in TDW. Image capture is unavoidably slow.

See the Benchmark document for the test machine's system info.

1: Image capture

Performance benchmark for Images output data using different render settings:

100 objects Pass masks Render quality Post-processing Screen size .png FPS
False ['_img'] 0 False 256 False 312
True ['_id'] 0 False 256 False 400
False ['_img'] 5 True 256 False 234
False ['_img'] 5 True 1024 False 44
True ['_id'] 0 False 1024 False 69
False ['_img'] 5 True 1024 True 41
True ['_img', '_id'] 5 True 1024 False 33
True ['_img', '_id'] 5 True 1024 True 30

2. Observation data

Performance benchmark to compare Images data to alternative observation data.

  • Render quality is 0
  • Post processing is disabled
  • Screen size is 256x256
  • Images are .jpg
  • There are 100 objects in the scene
  • Note that Occlusion is slow because it renders two images; it would normally require 2 frames to get the same data as Occlusion's 1 frame.
_img pass _id pass IdPassSegmentationColors Occlusion FPS
True False False False 292
False True False False 350
False False True False 288
False False False True 196

3. Image capture (all)

This is a simple benchmark of all image passes.

  • Render quality is 5
  • Post processing is enabled
  • Screen size is 256x256
  • The _img pass is a .png
  • There are 15 objects in the scene

Result: 45 FPS

How to run TDW's image capture performance benchmarks

  1. Follow instructions in the Benchmark document for cloning the repo, downloading the build, etc.
  2. cd path/to/tdw/Python/benchmarking (replace path/to with the actual path)
  3. python3 image_capture.py or python3 observation_data.py or python3 image_capture_all.py
  4. Run the build
  5. Wait for the performance benchmark to complete (this might take up to five minutes).
  6. Compare your results to those listed above

Next: Object data

Return to the README