Multiview Orthographic Feature Transformation for 3D Object Detection
Multiview 3D object detection on MultiviewC dataset through moft3d.
We propose a novel method, MOFT3D, for multiview 3D object detection and MultiviewC, a synthetic dataset, for multi-view detection in occlusion scenarios.
- MultiviewC dataset
- MOFT3D Code
The MultiviewC dataset mainly contributes to multiview cattle action recognition, 3D objection detection and tracking. We build a novel synthetic dataset MultiviewC through UE4 based on real cattle video dataset which is offered by CISRO.
The MultiviewC dataset is generated on a 37.5 meter by 37.5 meter square field. It contains 7 cameras monitoring cattle activities. The images in MultiviewC are of high resolution, 1280×720 and synthetic animals in our dataset are highly realistic.
- download (from Baidu Drive
Extraction Code: 6666and copy the
calibrationsfolder into this repo.
Build your own version
Please refer to this repo for MultiviewC dataset toolkits.
This repo is contributed to the code for MOFT3D.
Training and Inference
Download the latest training documents to
~/experiments folder from BaiduDrive
pwd:6666 or GoogleDrive and unzip them. This training documents contains the checkpoints of model, optimizer and scheduler and tensorboard containing the training details. Notice, this is not the final released version of MOFT3D.
There are two metrics to evaluate the performance of model. MODA, MODP, Precission and Recall are used to evaluate detection performance such as the detection in occlusion scenes. These metrics need to successfully run in matlab environment. Please refer to here for more details.
Even though, the python implementation of these metrics mentioned above is also provided, it need to select the distance threshould to detemine to positive samples，which is not objective enough. Thus, it is recommended to select the official implementation of matlab.
When it comes to the AP, AOS, OS metrics, we need to install cuda environment and build the toolkit for 3D rotated IoUs calculation. Please refer to this repo for more details.