[Paper Review] Concealed Object Detection

Deng-Ping Fan, Ge-Peng Ji+2 more|arXiv (Cornell University)|Feb 20, 2021

Adversarial Robustness in Machine Learning123 references4 citations

TL;DR

This paper introduces the first comprehensive study on concealed object detection (COD), proposing the large-scale COD10K dataset with 10,000 images and rich annotations across 78 categories. It presents SINet, a simple yet effective search-and-identification network that achieves state-of-the-art performance on all benchmark COD datasets, outperforming 12 advanced baselines without complex components.

ABSTRACT

We present the first systematic study on concealed object detection (COD), which aims to identify objects that are "perfectly" embedded in their background. The high intrinsic similarities between the concealed objects and their background make COD far more challenging than traditional object detection/segmentation. To better understand this task, we collect a large-scale dataset, called COD10K, which consists of 10,000 images covering concealed objects in diverse real-world scenarios from 78 object categories. Further, we provide rich annotations including object categories, object boundaries, challenging attributes, object-level labels, and instance-level annotations. Our COD10K is the largest COD dataset to date, with the richest annotations, which enables comprehensive concealed object understanding and can even be used to help progress several other vision tasks, such as detection, segmentation, classification, etc. Motivated by how animals hunt in the wild, we also design a simple but strong baseline for COD, termed the Search Identification Network (SINet). Without any bells and whistles, SINet outperforms 12 cutting-edge baselines on all datasets tested, making them robust, general architectures that could serve as catalysts for future research in COD. Finally, we provide some interesting findings and highlight several potential applications and future directions. To spark research in this new field, our code, dataset, and online demo are available on our project page: http://mmcheng.net/cod.

Motivation & Objective

• To address the lack of large-scale, densely annotated datasets for concealed object detection (COD), a critical yet underexplored task in computer vision.
• To establish a standardized benchmark for COD by introducing a new dataset with rich annotations, including categories, bounding boxes, instance masks, attributes, and matting-level labels.
• To develop a robust, generalizable deep learning framework for COD that outperforms existing state-of-the-art methods.
• To explore the feasibility of using animal-inspired search-and-identification strategies for detecting visually camouflaged objects.
• To stimulate future research by identifying ten open directions, including multi-modal detection, universal networks, and neural architecture search.

Proposed method

• Proposes COD10K, a large-scale dataset of 10,000 real-world images with 78 object categories, annotated with category, bounding box, object-level mask, instance-level mask, edge maps, and challenging attributes (e.g., shape complexity, occlusion, indefinable boundaries).
• Introduces SINet, a simple but powerful end-to-end network that adopts a search-and-identification strategy inspired by animal foraging behavior, using a dual-branch architecture to first locate potential regions and then refine the segmentation.
• Employs a multi-level supervision strategy with auxiliary supervision at different stages to enhance feature learning and improve mask prediction accuracy.
• Uses a combination of cross-entropy and Dice loss for training, enabling better optimization for dense prediction tasks like instance segmentation.
• Applies a self-supervised pre-training strategy on a subset of COD10K to improve generalization, especially in low-resource scenarios.
• Validates the framework on multiple existing COD benchmarks, demonstrating strong generalization and robustness across diverse real-world scenarios.

Experimental results

Research questions

• RQ1 How can a large-scale, densely annotated dataset for concealed object detection be constructed to support comprehensive benchmarking and model evaluation?
• RQ2 Can a simple, end-to-end deep learning framework outperform complex state-of-the-art models in concealed object detection without architectural bells and whistles?
• RQ3 To what extent does a search-and-identification strategy, inspired by animal hunting behavior, improve detection performance on highly camouflaged objects?
• RQ4 What are the key attributes that make concealed object detection particularly challenging, and how do they affect model performance?
• RQ5 What are the most promising future research directions for advancing concealed object detection beyond current limitations?

Key findings

• SINet achieves state-of-the-art performance on all tested COD datasets, outperforming 12 cutting-edge baselines across all evaluation metrics, including F-measure, mean Dice, and mean IoU.
• The search-and-identification strategy in SINet proves highly effective, demonstrating that a two-stage approach—first searching for candidate regions and then identifying precise boundaries—yields superior results compared to one-stage detection.
• The COD10K dataset contains 10,000 images with rich annotations, including matting-level labels that take approximately 60 minutes per image to produce, ensuring high-quality supervision.
• The dataset includes 78 diverse object categories across terrestrial, amphibian, flying, and aquatic types, with challenging attributes such as shape complexity, occlusion, and indefinable boundaries.
• SINet achieves this performance with only 4 hours of training time, indicating high efficiency and practicality for real-world deployment.
• The study reveals that current salient object detection models generalize poorly to concealed objects, highlighting the need for dedicated COD frameworks and datasets.