Jiangsan Zhao

Research Scientist

(+47) 413 23 321
jiangsan.zhao@nibio.no

Place
Apelsvoll

Visiting address
Nylinna 226, 2849 Kapp

Abstract

Neural Radiance Fields (NeRF) have been widely adopted for reconstructing high-quality 3D scenes from 2D RGB images. However, achieving accurate 3D object segmentation within these reconstructed scenes remains challenging. Existing NeRF-based segmentation methods either rely on post-processing (SA3D), which produces noisy point clouds due to the absence of density field optimization, or employ joint training with additional segmentation heads (FruitNeRF), which can lead to suboptimal performance due to conflicting learning objectives. In this work, we propose InvNeRF-Seg (Input-substitution NeRF for Segmentation), a two-stage fine-tuning strategy for 3D object segmentation that preserves the original NeRF architecture and loss function entirely. We first train a standard NeRF on RGB images and then fine-tune it using 2D segmentation masks formatted as RGB-like inputs, without introducing any architectural modifications or additional loss functions. This input-substitution approach reshapes the density field to align with object regions while suppressing background density. We validate InvNeRF-Seg through comprehensive ablation studies examining the roles of density and color MLPs, loss function choices, and training strategies. Field density analysis reveals consistent semantic refinement: densities of object regions increase while background densities are suppressed. Experiments on synthetic fruit datasets and real-world soybean imagery demonstrate that InvNeRF-Seg produces cleaner 3D segmented point clouds compared to both SA3D and FruitNeRF, enabling more accurate downstream object counting. The method is further validated on a self-collected soybean dataset to demonstrate its applicability in real-world agricultural scenarios.

To document

Abstract

In broiler breeding, precise counting is crucial for improving production efficiency and ensuring animal welfare. Nevertheless, counting chickens precisely is a challenging task especially when young chicks always huddle for warmth. Although deep learning has been widely taken in different counting related tasks, more accurate localization and counting of chickens in high stocking density scenes still has not been well investigated. We propose a point supervised dense chickens flock counting network (PCCNet), which directly utilizes points as learning targets. The network adopts information feature fusion to assist the identification of broilers high stocking density scenes. In addition, considering the distance of neighboring points as matching cost in point matching algorithms is advantageous for generating more reasonable matching results, facilitating model convergence. To validate the effectiveness of the proposed network, a Chicken Counting Dataset (CCD) is built, consisting of two subsets separated by different ages: CCD_A and CCD_B. The accuracies of PCCNet on the two subsets of CCD are 97.85% and 97.06%, with corresponding Mean Absolute Errors (MAE) of 1.966 and 5.173, and Root Mean Square Errors (RMSE) values of 3.474 and 7.034, respectively. Our model achieves better broiler counting performance than other state-of-the-art (SOTA) methods.