ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

Evaluation of Mask2Former with ConvNeXt V2 Backbones for On-Edge Instance Segmentation in Poultry Images

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2601509.(doi:10.13031/aim.202601509)
Authors:   Aashish Poudel, Sudhendu R. Sharma, Jeremiah D. Davis, Joseph L. Purswell, Tami M. Brown-Brandl
Keywords:   edge AI, instance segmentation, Jetson Orin Nano, Mask2Former, poultry monitoring, TensorRT.

Abstract. Precision livestock farming in poultry production is challenged by large flock sizes, limiting the feasibility of individual-level monitoring. Vision-based methods offer a scalable solution, but practical deployment requires efficient on-edge processing. Instance segmentation enables extraction of individual bird information from dense group settings while supporting on-edge estimation of flock activity. The objective of this study was to evaluate Mask2Former with ConvNeXt V2 backbones for accurate and efficient on-edge instance segmentation in poultry images. Five-fold cross-validation was conducted on datasets collected under artificial (AL) and natural (NL) lighting, with cross-dataset evaluation used to assess generalizability. A vanilla Mask2Former with a Swin-Tiny backbone served as the primary baseline, while YOLO-based models were included for runtime context. Models were converted to TensorRT FP16 and FP32 engines and benchmarked on an NVIDIA Jetson Orin Nano Super. Among ConvNeXt V2 backbones, femto achieved the highest segmentation accuracy (74.37% AP50:95 on AL; 72.70% on NL). Under FP16, vanilla Mask2Former and ConvNeXt V2 atto showed comparable deployment efficiency (6.55 vs 6.91 W), while atto provided the smallest engine footprint (51.8 MB). Cross-dataset evaluation demonstrated asymmetric transfer, with NL-trained models generalizing better to AL images. FP16 improved inference speed and power efficiency across all variants. Notably, lower computational cost did not directly translate to faster inference on edge hardware. Although YOLO-based models achieved higher inference speeds, Mask2Former produced higher-quality masks for downstream tasks such as activity and space-use estimation. ConvNeXt V2 femto is recommended when accuracy is the priority, while atto is better suited for memory-constrained edge deployment, with the Swin-Tiny baseline remaining competitive in power efficiency.

(Download PDF)    (Export to EndNotes)