Commits

Breadcrumbs

History for

ononnx

User selector

All users

Datepicker

All time

Commit History

Commits on Dec 9, 2020

• working onnx trt engine

  simutisernestas

  committed
  e0cf2a3

  Copy full SHA for e0cf2a3

Commits on Apr 30, 2019

• Actually implemented the has_gt option. For whatever reason it didn't do anything before.

  dbolya

  committed
  a4933c0

  Copy full SHA for a4933c0

Commits on Apr 25, 2019

• Removed some unnecessary functions.

  dbolya

  committed
  3d3c375

  Copy full SHA for 3d3c375

Commits on Mar 30, 2019

• Some tweaks

  dbolya

  committed
  8977d62

  Copy full SHA for 8977d62

• Changed codec to mp4v

  dbolya

  committed
  1481b72

  Copy full SHA for 1481b72

• Added the ability to save video outputs.

  dbolya

  committed
  c239148

  Copy full SHA for c239148

Commits on Mar 29, 2019

• Fixed no detection results for prep_display.

  dbolya

  committed
  bc67fd4

  Copy full SHA for bc67fd4

Commits on Mar 22, 2019

• Added a --detect setting to run YOLACT as a detector.

  dbolya

  committed
  452bc95

  Copy full SHA for 452bc95

Commits on Mar 21, 2019

• Fixed the video doohickey and made --score_threshold faster.

  dbolya

  committed
  b94b4a1

  Copy full SHA for b94b4a1

• Traditional NMS implemented!

  dbolya

  committed
  5cbdcc3

  Copy full SHA for 5cbdcc3

• Added missing timers from the quantitative eval branch when outputting a COCO JSON file.

  dbolya

  committed
  e82a7cb

  Copy full SHA for e82a7cb

Commits on Mar 19, 2019

• Implemented Chong's vectorized cropping and refactored a bunch of the postprocessing code to accommodate it better. As a side effect, not only it the model 3+ms faster, but it's also more accurate!…

  Show description for 89fabd0
  dbolya

  committed
  89fabd0

  Copy full SHA for 89fabd0

Commits on Mar 17, 2019

• That was bad?

  dbolya

  committed
  2b6a59c

  Copy full SHA for 2b6a59c

Commits on Mar 14, 2019

• Fixed the BRG issue when using --image mode, added the ability to write '--image:x.png:y.png' to save an output instead of displaying. Also changed -> to : because the former required you to wrap t…

  Show description for 480bb56
  dbolya

  committed
  480bb56

  Copy full SHA for 480bb56

Commits on Mar 12, 2019

• Save images as pngs instead.

  dbolya

  committed
  50f3a53

  Copy full SHA for 50f3a53

• Idk are these good colors?

  dbolya

  committed
  708dcc0

  Copy full SHA for 708dcc0

• Added the --images argument that allows you to evaluate on a whole folder of images at once.

  dbolya

  committed
  23b0415

  Copy full SHA for 23b0415

• Made the opencv rendering prettier.

  dbolya

  committed
  fdaa398

  Copy full SHA for fdaa398

Commits on Mar 2, 2019

• Added a cuda synchronize to the benchmark. This shouldn't affect speed but I just wanted to make sure.

  dbolya

  committed
  9bb12fe

  Copy full SHA for 9bb12fe

Commits on Feb 24, 2019

• Sped up evalvideo considerably to the point where video_multiframe actually does something. Gotta go fast.

  dbolya

  committed
  67fc5d8

  Copy full SHA for 67fc5d8

Commits on Feb 22, 2019

• Refactored evalvideo and added an attempt at pushing through multiple frames at once. However, it looks like we're too CPU-bound at the moment, so this doesn't end up being faster. I'll take a look…

  Show description for 48e2478
  dbolya

  committed
  48e2478

  Copy full SHA for 48e2478

Commits on Feb 20, 2019

• Added an option to not display text on the displayed masks for maximum immersion.

  dbolya

  committed
  5e9820c

  Copy full SHA for 5e9820c

Commits on Feb 18, 2019

• Changed evaluating video to use a consistent color per class so no flickering detection colors. Also fixed an issue where some variables in postprocessing required grad for some reason (likely beca…

  Show description for d1c6234
  dbolya

  committed
  d1c6234

  Copy full SHA for d1c6234

Commits on Feb 17, 2019

• Added the ability to evaluate a video in real-time with little overhead. Also considerably sped up prep_display by applying masks to the image on the GPU.

  dbolya

  committed
  b80b274

  Copy full SHA for b80b274

Commits on Feb 10, 2019

• Did a fair bit of refactoring and changed the way we load datasets. I've included and this enables use to evaluate on the test set.

  dbolya

  committed
  d38bad1

  Copy full SHA for d38bad1

• Added the ability to specify the desired dataset in commandline.

  dbolya

  committed
  3340868

  Copy full SHA for 3340868

Commits on Jan 24, 2019

• Fixed a bug where crowd annotations weren't being used in eval, making us produce a different mAP than COCOEval. Idk how this got here :/

  dbolya

  committed
  a94f1db

  Copy full SHA for a94f1db

Commits on Jan 8, 2019

• Added the ability to test loose images by specifying the image path with --image=myimage.png. Also added the ability to threshold scores with --score_threshold=min_val.

  dbolya

  committed
  0ad4eae

  Copy full SHA for 0ad4eae

Commits on Nov 12, 2018

• Added an eval argument for fast box NMS.

  dbolya

  committed
  4367675

  Copy full SHA for 4367675

Commits on Oct 9, 2018

• Fixed a bug in eval.py that caused a crash with --display when no detections are output.

  dbolya

  committed
  b627472

  Copy full SHA for b627472

• Fixed eval.py not using the dedicated function to load waits (it didn't matter until now).

  dbolya

  committed
  e55b9b0

  Copy full SHA for e55b9b0

Commits on Oct 2, 2018

• Finally added the ability to set a crowd iou threshold to count negatives that have a higher crowd iou than that to be considered neutral. I added 2 configs to test it and whether or not I borked a…

  Show description for 367c3e2
  dbolya

  committed
  367c3e2

  Copy full SHA for 367c3e2

Commits on Sep 28, 2018

• Oh... I figured out what the problem was. If you use --resume=latest and no latest file is found, the network starts from iteration 0, as intended. However, the way I implemented it, that branch ac…

  Show description for 898babd
  dbolya

  committed
  898babd

  Copy full SHA for 898babd

Commits on Sep 27, 2018

• Added the ability to specify --no_crop to eval.py to not crop the final masks.

  dbolya

  committed
  a4c4d82

  Copy full SHA for a4c4d82

Commits on Sep 26, 2018

• Added --trained_model=latest functionality to eval.py.

  dbolya

  committed
  ff9fe1f

  Copy full SHA for ff9fe1f

Pagination

Previous

Next