Research map / Detection and segmentation / CNN features
Depth-Aware Concealed Crop Detection in Dense Agricultural Scenes
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Concealed Crop Detection (CCD), an RGB-D agricultural concealed-object dataset ACOD-12K, and RISNet, a PVT-based depth-guided, multi-iteration segmentation network that reports state-of-the-art on CCD and classic COD benchmarks.
Problem
Concealed object detection methods focus on animals or humans and fail on dense, small, occluded crops in agricultural scenes, where RGB alone gives insufficient information.
Approach
Combine a PVT encoder with ASPP multi-scale receptive fields, fuse RGB and depth features through element-wise add/multiply plus dual attention, and iteratively refine the coarse prediction used as gate attention to improve small-object detection.
Key results
- RISNet achieves S-alpha 0.866 and weighted F-measure 0.803 on ACOD-12K, surpassing HitNet (0.853 / 0.787) by about 1.45% on average.
- On ACOD-12K, weighted F-measure exceeds the second-best RGB-D method XMSNet by 4.9%.
- On COD10K, RISNet reaches S-alpha 0.873 and weighted F-measure 0.799, and on NC4K 0.882 and 0.834 (trained on COD10K+CAMO).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use RISNet instead of HitNet | CCD · ACOD-12K | Sα: 0.853 -> 0.866 | Helped |
| use RISNet instead of HitNet | CCD · ACOD-12K | Fωβ: 0.787 -> 0.803 | Helped |
| use RISNet instead of HitNet | CCD · ACOD-12K | Eθ: 0.955 -> 0.967 | Helped |
| use RISNet instead of DaCOD | COD detection · CAMO | Fωβ: 0.796 -> 0.827 | Helped |
| use RISNet instead of HitNet | COD detection · COD10K | Eθ: 0.932 -> 0.931 | Hurt |
| remove multi-scale feature encoding (replace PVT with Res2Net-50 and remove ASPP) (ablation) | RISNet · ACOD-12K | Sα: 0.866 -> 0.855 | Hurt |
| remove designed depth-guided feature decoding and use direct modality concatenation (ablation) | RISNet · ACOD-12K | Fωβ: 0.803 -> 0.790 | Hurt |
| remove iterative feature refinement (ablation) | RISNet · ACOD-12K | Fωβ: 0.803 -> 0.785 | Hurt |
| increase iterations from 1 to 3 (ablation) | RISNet · ACOD-12K | Fωβ: 0.792 -> 0.803 | Helped |
| increase iterations from 3 to 4 (ablation) | RISNet · ACOD-12K | Fωβ: 0.803 -> 0.802 | Hurt |
Surprising findings
- Single-modal COD methods such as FSPNet score lower on ACOD-12K (S-alpha 0.719) than on classic COD benchmarks, indicating the agricultural domain is much harder for existing COD models.
Limitations
- Mean absolute error metric excluded from CCD evaluation, so the benchmark uses only three metrics.
- Dataset covers ten crop categories collected from specific orchards and farms, limiting diversity.
- Iterative refinement requires multiple passes (3 iterations), adding inference cost.
Experimental setup
Datasets: ACOD-12K COD10K CAMO
Benchmarks: ACOD-12K CAMO COD10K NC4K
Metrics: S-alpha weighted F-measure E-measure
Baselines: SINet SINet-V2 ZoomNet PopNet HitNet FSPNet DaCOD XMSNet FEDER SegMaR
Builds on: HitNet PVT ASPP SINet ZoomNet
Closest papers by meaning
- Can Large Multimodal Models Understand Agricultural Scenes? Benchmarking with AgroMind (NeurIPS 2025 · similarity 0.84)
- Combining Deep Learning and Street View Imagery to Map Smallholder Crop Types (AAAI 2024 · similarity 0.83)
- Fields of The World: A Machine Learning Benchmark Dataset for Global Agricultural Field Boundary Segmentation (AAAI 2025 · similarity 0.83)
- GTPBD: A Fine-Grained Global Terraced Parcel and Boundary Dataset (NeurIPS 2025 · similarity 0.82)
- Cropper: Vision-Language Model for Image Cropping through In-Context Learning (CVPR 2025 · similarity 0.82)
- Exact: Exploring Space-Time Perceptive Clues for Weakly Supervised Satellite Image Time Series Semantic Segmentation (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.