# Two-Stream Bidirectional Interaction Network Based on RGB-D Images for Duck Weight Estimation

## TL;DR

This paper introduces a non-contact method using RGB-D images and a two-stream network to accurately estimate duck weight, reducing stress and supporting precision poultry management.

## Contribution

A novel two-stream bidirectional interaction network using RGB-D images for accurate duck weight estimation is proposed.

## Key findings

- The method achieves the lowest MAE of 0.1550 on a new RGB-D duck image dataset.
- The approach outperforms existing methods in duck weight estimation accuracy.
- The method supports automated data collection and promotes animal welfare in poultry management.

## Abstract

Non-contact weight measurement in ducks is critical for reducing duck stress and enabling precision poultry management. We propose an automated method using RGB-D images and a two-stream bidirectional network to estimate duck weight accurately. The method leverages separate branches in the encoder to extract RGB and depth features, complemented by cross-modal fusion to enhance feature complementation between modalities. The decoder then combines multi-scale features for weight regression. Validated on a novel dataset of 2865 RGB-D duck images, our approach achieves the lowest MAE of 0.1550, outperforming existing methods. This technology eliminates physical handling stress, automates growth data collection, and supports informed decisions for precision feeding and health monitoring. By promoting animal welfare, it advances sustainable agricultural practices in the poultry industry.

An automated non-contact weight measurement method for ducks is beneficial for preventing the stress response of ducks and, thus, promoting their healthy development. We propose a two-stream bidirectional interaction network that depends on RGB-D pictures to accurately determine the weight of ducks. We developed two-stream branches in the encoder to extract texture appearance information and spatial structure information from RGB images and depth images, respectively. Besides, we employed a cross-modality feature supplement module in the encoder to facilitate mutual learning and complementarity between these two modalities. Finally, a decoder is designed to combine the multi-scale characteristics of these two modalities and feed the fused features into the regression module to determine the final weight of the duck. For the experimental analysis of this study, we built a new dataset of RGB-D duck images consisting of 2865 pairs of RGB-D images captured from the bird-eye view. The comparative experimental results show that the proposed method could effectively estimate the duck weight with an MAE of only 0.1550, outperforming all the comparison methods on this dataset. This automated, non-contact weight measurement method can eliminate stress responses caused by human intervention. This method enables the automated collection of growth data, supporting precision feeding and health management decisions. It drives the digital and welfare-oriented transformation of the livestock industry, enhancing production efficiency while promoting animal welfare and sustainable agricultural practices.