FruitLangGS is a language-guided 3D fruit counting framework that uses an adaptive dense Gaussian splatting pipeline with radius-aware pruning and tile-based rasterization to reconstruct orchard-scale scenes. Unlike existing pipelines that rely on multi-view 2D segmentation and dense volume sampling, FruitLangGS filters the compressed CLIP-aligned semantic vectors contained within each Gaussian through a double-threshold cosine similarity mechanism to retrieve Gaussians relevant to the target prompt without retraining or image-space masks, suppressing common distractors (e.g., leaves). The selected Gaussians are sampled from a dense point cloud and geometrically clustered to estimate fruit instances, and are robust to severe occlusion and viewpoint variations. Experiments on nine different orchard-scale datasets demonstrate that FruitLangGS consistently outperforms existing pipelines in instance counting recall, avoids multi-view segmentation fusion errors, and achieves up to 99.7% recall on the Pfuji-Size_Orch2018 orchard dataset. Additional ablation studies confirm that language-conditional semantic embeddings and double-threshold prompt filtering are essential for suppressing distractors and improving counting accuracy under severe occlusion. Beyond fruit counting, the same framework enables prompt-based 3D semantic retrieval without retraining, highlighting the potential of language-guided 3D recognition for scalable agricultural scene understanding.
