Abstract
Concealed Crop Detection (CCD) addresses the challenge of identifying small, densely packed agricultural products that blend into their surroundings. This paper introduces ACOD-12K, the first large-scale RGB-D dataset for CCD, and RISNet, a Recurrent Iterative Segmentation Network. By leveraging depth information and an iterative refinement process, RISNet effectively distinguishes crops from overlapping foliage, setting a new benchmark for agricultural object detection.
Introduction: The Challenge of Concealed Crops
Traditional Concealed Object Detection (COD) typically focuses on animals or humans using camouflage to evade detection. In agriculture, however, crops like peppers, tomatoes, and cucumbers exhibit "passive" concealment. They are often the same color as the surrounding leaves, extremely small, and buried under dense, overlapping foliage. Standard RGB-based models struggle in these high-density environments because the visual signal of the fruit is nearly indistinguishable from the background noise.
Methods: Recurrent Iterative Segmentation Network (RISNet)
The proposed RISNet architecture addresses these challenges through three primary components:
- Concealed Feature Encoder (CFE): Utilizing a Pyramid Vision Transformer (PVT) and Atrous Spatial Pyramid Pooling (ASPP), the encoder captures multi-scale features, allowing the model to detect crops of varying sizes and growth stages.
- Depth-Guided Feature Decoder (DFD): This module integrates depth maps to provide a geometric "signature" for crops. Using Multi-modal Feature Fusion (MFF), it combines RGB and depth data through element-wise operations to identify the physical protrusions of fruits that color alone cannot reveal.
- Iterative Feature Refinement (IFR): The model employs a three-stage iterative process. Initial predictions are fed back into the network as "gate attention," focusing computational resources on refining the boundaries of small or highly occluded objects.
The ACOD-12K Dataset
To support research in CCD, the authors released ACOD-12K, a comprehensive multi-modal dataset:
- Scale and Resolution: Contains 6,092 high-resolution (1080p) images.
- Multi-modality: The first large-scale RGB-D dataset specifically for concealed object detection.
- Diversity: Features 10 crop categories, including watermelon, bean, tomato, pepper, kiwi fruit, cucumber, zucchini, peach, loquat, and plum.
- Complexity: Images average 11 objects each, representing significantly higher density and occlusion levels than standard COD datasets.
Key Insights: Depth as a Geometric Signature
In dense agricultural scenes, chromatic contrast is often near zero. The primary insight of this work is that depth information acts as a critical tie-breaker. While a green cucumber may look like a leaf, its 3D shape and protrusion provide a distinct geometric signature. RISNet’s ability to fuse these modalities allows it to "break" the passive camouflage of the crops. Furthermore, the iterative refinement process mimics human visual attention, gradually narrowing focus from a general area to precise object boundaries.
Results and Performance
RISNet was evaluated against 19 state-of-the-art methods. On the CCD task, it achieved a structure measure ($S_\alpha$) of 0.866, outperforming the nearest competitor, HitNet, by 1.45%. Qualitative analysis demonstrates that RISNet produces cleaner, more distinct masks than other models, even when crops are 80-90% occluded. The model also showed strong performance on traditional animal-based datasets (CAMO, COD10K), proving its generalizability beyond agriculture.
Limitations and Future Outlook
Despite its success, RISNet faces challenges regarding computational overhead due to its iterative transformer-based architecture, which may limit deployment on low-power edge devices like agricultural drones. Additionally, the model's performance is tied to the quality of depth maps, which can be affected by environmental factors like harsh sunlight or rain. Future research may focus on model pruning for real-time applications and zero-shot learning to expand detection capabilities to a wider variety of global crops.
Summary updated