Skip to content

Latest commit

 

History

History
130 lines (97 loc) · 6.83 KB

RESULTS.md

File metadata and controls

130 lines (97 loc) · 6.83 KB

Results

  1. Cube R-CNN on Omni3D
  2. Comparison with Competing Methods
  3. Cube R-CNN on KITTI test

Cube R-CNN on Omni3D

The Cube R-CNN evaluation produces two tables which summarize performance on the evaluation set. The first is a performance analysis table and the second is the main Omni3D performance table.

Below are the results of the DLA34 Cube R-CNN model trained on the full Omni3D training set and evaluated on the test set.

  1. Performance Analysis Table
Dataset #iters AP2D AP3D AP3D@15 AP3D@25 AP3D@50 AP3D-N AP3D-M AP3D-F
SUNRGBD_test final 16.0053 15.3321 21.6640 16.9483 5.13407 15.3321 nan nan
Hypersim_test final 12.2447 7.47746 10.0416 7.58102 2.25935 7.95103 0.586047 0
ARKitScenes_test final 41.3007 41.7261 53.0945 45.4191 19.2622 41.7267 0 nan
Objectron_test final 56.4603 50.8374 65.6977 54.0398 22.4584 50.8374 nan nan
KITTI_test final 41.3125 32.5909 41.9073 34.5554 16.236 56.6344 36.0952 16.4902
nuScenes_test final 36.31 30.059 39.1756 32.1927 14.5962 47.4605 34.8069 11.9634
Concat final 27.6387 23.266 30.8423 24.865 9.51163 27.9432 12.0738 8.49733
  1. Omni3D Performance Table -- To be used to compare with Cube R-CNN
Dataset #iters AP2D AP3D
SUNRGBD_test final 16.0053 15.3321
Hypersim_test final 12.2447 7.47746
ARKitScenes_test final 41.3007 41.7261
Objectron_test final 56.4603 50.8374
KITTI_test final 41.3125 32.5909
nuScenes_test final 36.31 30.059
Omni3D_Out final 38.8662 33.0019
Omni3D_In final 23.4123 20.0325
Omni3D final 27.6387 23.266

The Omni3D entry (last row) gives performance on the full Omni3D test set. This is what we report in our paper and what should be used to compare to Cube R-CNN. The tables also report performance on the outdoor and indoor subsets of the test set, in the Omni3D_Out and Omni3D_In entries respectively.

Performance on Omni3D Indoor and Outdoor

Below we provide Cube R-CNN performance when trained and evaluated on the indoor (Omni3D_In) and outdoor (Omni3D_Out) splits.

Omni3D_In

Here we train and evaluate Cube R-CNN on the Omni3D_In split which consists of {Hypersim, SUN RGB-D, ARKitScenes}.

  1. Performance Analysis Table
Dataset #iters AP2D AP3D AP3D@15 AP3D@25 AP3D@50 AP3D-N AP3D-M AP3D-F
SUNRGBD_test final 18.1246 16.7919 23.7455 18.8856 5.3282 16.792 nan nan
Hypersim_test final 13.3741 7.30641 9.75961 7.26068 2.61309 7.81662 0.690567 0
ARKitScenes_test final 43.7697 43.5845 56.7073 47.2705 18.9982 43.5858 0 nan
Concat final 19.2801 15.0396 20.4719 16.1827 5.51773 15.5207 0.66831 0
  1. Omni3D Performance Table
Dataset #iters AP2D AP3D
SUNRGBD_test final 18.1246 16.7919
Hypersim_test final 13.3741 7.30641
ARKitScenes_test final 43.7697 43.5845
Omni3D_Out final nan nan
Omni3D_In final 19.2801 15.0396
Omni3D final nan nan
Omni3D_Out

Here we train and evaluate Cube R-CNN on the Omni3D_Out split which consists of {KITTI, nuScenes}.

  1. Performance Analysis Table
Dataset #iters AP2D AP3D AP3D@15 AP3D@25 AP3D@50 AP3D-N AP3D-M AP3D-F
nuScenes_test final 38.3153 32.6165 41.1472 33.8193 17.7208 43.6913 37.7967 15.0941
KITTI_test final 43.7112 35.9988 44.8352 37.5502 19.7639 52.0115 40.0687 16.3201
Concat final 39.1041 31.8352 40.2945 33.0542 16.809 45.3547 37.2141 13.9689
  1. Omni3D Performance Table
Dataset #iters AP2D AP3D
nuScenes_test final 38.3153 32.6165
KITTI_test final 43.7112 35.9988
Omni3D_Out final 39.1041 31.8352
Omni3D_In final nan nan
Omni3D final nan nan

Comparison with Competing Methods on Omni3D

We compare Cube R-CNN to recent state-of-the-art 3D object detection methods SMOKE, FCOS3D, PGD and ImVoxelNet. We augment the first three with our virtual camera (vc) feature. ImVoxelNet by design uses camera intrinsics to unproject to 3D.

Method AP3D
ImVoxelNet 9.4
SMOKE 9.6
SMOKE + vc 10.4
FCOS3D 9.8
FCOS3D + vc 10.6
PGD 11.2
PGD + vc 15.4
Cube R-CNN 23.3

Cube R-CNN Performance KITTI test

We evaluate and compare Cube R-CNN on KITTI test using KITTI's evaluation server which reports AP3D at a 70% 3D IoU. Note that Cube R-CNN is not tuned for the KITTI benchmark. We train Cube R-CNN on KITTI only and do not perform any data augmentations or model ensembling at test time.

Method Easy Med Hard
SMOKE 14.03 9.76 7.84
ImVoxelNet 17.15 10.97 9.15
PGD 19.05 11.76 9.39
GUPNet 22.26 15.02 13.12
Cube R-CNN 23.59 15.01 12.56