Optimization on multi-object tracking and segmentation in pigs' weight measurement

Weight of pigs is highly correlated to their health. At present, 3D cameras can get spatial information, which develop non-contacting weight measurement. Separating pigs from the background is the first step, and tracking in a short video can make the weight more accurate than predicting weight on single image. Multi-Object Tracking and Segmentation (MOTS) in a video has received more attention with adding association embedding branch into instance segmentation network. Despite its success, the MOTS network has a crucial problem in practical application, that the predicted masks do not fit the objects well. The reason is low resolution of the feature maps in mask branch. So we improve the mask generation branch by cascading deconvolution layer and atrous convolution layer flexibly. The experimental results show that two deconvolution layers cooperating with two atrous convolution layers perform better. In pigs' weight measurement, this method outputs more precise masks than original network.

Automated summary

By flexibly cascading deconvolution layers and atrous convolution layers, we improved the mask generation branch, resulting in more precise weight measurement of pigs.