Home

Awesome

News

BOP19_CDPN_2019ICCV

The modified version of CDPN ("CDPN: Coordinates-Based Disentangled Pose Network for Real-Time RGB-Based 6-DoF Object Pose Estimation" on ICCV 2019) for BOP: Benchmark for 6D Object Pose Estimation Challenge 2019.

Our test environments

Detection

For detection, we trained a RetinaNet for each dataset on mmdetection.

Pose Estimation

In the BOP 2019 challenge, different from the paper, both of the rotation and translation are solved from the built 2D-3D correspondences by PnP algorithm. We trained a CDPN model for each object.

Data Preparation

  1. Download the 7 core datasets from the BOP website

  2. Download our trained models and detection results.

  3. Prepare the data as follows:

    Note:

    • models_eval: downloaded official models;
    • test/test_primesense: downloaded official BOP19 test set;
    • val:optionally, downloaded official val set;
    • trained_models: our provided trained models;
    • bbox_retinanet: our provided detection results;
    • exp: save the test result files
Root
├── dataset
│   ├── lmo_bop19
│   │   ├── models_eval 
│   │   └── test 
│   ├── tudl_bop19
│   │   ├── models_eval 
│   │   └── test 
│   ├── hb_bop19
│   │   ├── models_eval
│   │   ├── val 
│   │   └── test
│   ├── icbin_bop19
│   │   ├── models_eval
│   │   └── test 
│   ├── itodd_bop19
│   │   ├── models_eval 
│   │   ├── val
│   │   └── test
│   ├── tless_bop19
│   │   ├── models_eval
│   │   └── test_primesense 
│   └── ycbv_bop19
│       ├── models_eval 
│       └── test
├── trained_models
│   ├── lmo
│   │   ├── obj_ape.checkpoint
│   │   └── ...
│   └── ...
├── bbox_retinanet
│   ├── lmo
│   │   ├── lmo_test_bop19_000002.json
│   │   └── ... 
│   └── ...
├── lib
├── tools
├── detection
└── exp

Run

  1. In 'tools' directory, run
  sh run.sh

It will first generate a .csv file to record the result of each object for each dataset. The final result files can be found in 'exp/final_result/CDPN_xxxx-test.csv'

  1. Use the Bop_toolkit for evaluation.