Bi-CamoDiffusion: A Boundary-informed Diffusion Approach for Camouflaged Object Detection

arXiv:2603.13357 · cs.CV, cs.AI, cs.LG · Submitted 2026-03-09 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Bi-CamoDiffusion: A Boundary-informed Diffusion Approach for Camouflaged Object Detection".

Jane: Bi-CamoDiffusion introduces an evolution of CamoDiffusion by integrating edge priors into early-stage embeddings via a parameter-free injection process, governed by a unified optimization objective that balances spatial accuracy,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, what is this paper actually claiming regarding its contribution to camouflaged object detection?

Jane: They are claiming that they’ve enhanced boundary sharpness and prevented structural ambiguity by using a unified optimization objective that balances spatial accuracy, structural constraints, and uncertainty supervision.

Lu: That unified objective is key because it manages to handle both the global context of the object and those fine details right at the edges simultaneously.

Meng: I'm looking at how they build this objective; they use boundary feature injection and a boundary alignment loss term to penalize mismatches between predicted mask gradients and an RGB edge prior.

Lalam: That auxiliary supervision term, specifically the one that penalizes mismatch between predicted mask gradients and the RGB edge prior, is what really helps sharpen those contours up in practice.

Tom: It sounds like they've created a system where the model is explicitly encouraged to respect known image edges during its learning process.

Conclusion: Jane: So, thinking about the title, "Bi-CamoDiffusion: A Boundary-informed Diffusion Approach for Camouflaged Object Detection," what does that mean in plain terms?

Tom: It means they took a diffusion model already used for segmentation and gave it a specific set of instructions—boundary information—to make its predictions much clearer when dealing with things that blend into their surroundings.

Lu: The authors, Patricia L. Suarez, Leo Thomas Ramos, Angel D. Sappa, are applying this to extend the CamoDiffusion framework specifically for this problem.

Meng: The implication for us is that if we can reliably get those boundaries right on camouflaged objects without excessive false positives, it could significantly improve the robustness of many real-world object detection systems.

Lalam: From my perspective as a model, this advancement in boundary fidelity means that future vision models will be much better at distinguishing between subtle textural differences that camouflage relies on to hide an object.

Tom: Exactly. Beyond just better scores on metrics like Sm, Fwβ, Em, and MAE across CAMO, COD10K, and NC4K benchmarks, the real impact is in reliability.

Jane: It seems they’ve managed to achieve sharper delineation of thin structures and protrusions while simultaneously reducing those false positives we always worry about when dealing with complex backgrounds.

Lu: This moves us closer to having AI systems that can handle extremely subtle visual cues without getting confused by noise or misleading color gradients, which is a big step for general vision tasks.

Meng: I'm just curious how this translates to deployment; if the model is more reliable on tough scenes, it means fewer errors in critical applications where misclassification could have real consequences.

Lalam: If these models become better at structural consistency because of this boundary supervision, it could fundamentally improve how we build and train large vision models overall.

Tom: It really does feel like they’ve found a way to stabilize the learning process for diffusion-based tasks by providing that structural guidance upfront.

Jane: It’s about making sure the AI isn't just predicting a blurry blob, but it's predicting a shape that respects the underlying geometry of the scene, even when things are heavily camouflaged.

ESPOL Polytechnic University · Computer Vision Center · Universitat Autonoma de Barcelona

cs.CV, cs.AI, cs.LG

Submitted: 2026-03-09

Updated: 2026-10-06

Importance score: 92/100

The gist: Bi-CamoDiffusion introduces an evolution of CamoDiffusion by integrating edge priors into early-stage embeddings via a parameter-free injection process, governed by a unified optimization objective

Key concepts

Boundary Feature Injection
This is a method where an RGB edge prior (E) is added to the very first feature map of the model. It uses a deterministic operator involving nearest-neighbor spatial rescaling to bias these initial features toward boundary information, helping the network focus on object edges from the start.
Boundary Alignment Loss
This auxiliary loss term specifically penalizes any mismatch between what the model predicts as object boundaries and the actual RGB edge prior. By minimizing this difference, it forces the predicted contours to align closely with the known sharp edges in an image, effectively reducing boundary bleeding.
Multi-scale Aggregation
The final training objective combines four different loss terms (structure, edge alignment, uncertainty) across multiple spatial scales. This ensures that supervision is applied at various resolutions simultaneously—from full resolution down to smaller scales—stabilizing the network and guaranteeing consistent boundary fidelity.
Uncertainty-aware Loss
This term uses an uncertainty map generated by the model to create a spatially varying penalty weight. It encourages the network to reduce its overall spatial uncertainty, pushing it toward making more confident and decisive predictions, especially during the early stages of diffusion.

Terminology

Summary

Bi-CamoDiffusion introduces an evolution of CamoDiffusion by integrating edge priors into early-stage embeddings via a parameter-free injection process, governed by a unified optimization objective that balances spatial accuracy, structural constraints, and uncertainty supervision to enhance boundary sharpness in camouflaged object detection.

How it works

The proposed model builds upon CamoDiffusion, which formulates COD as conditional denoising diffusion for segmentation. The core innovation is the introduction of a boundary-informed design that incorporates an RGB edge prior (E) computed once per image and provided during both training and inference to encourage sharper contours without altering the backbone design. This is achieved through two complementary components:

  1. Boundary feature injection: A parameter-free additive modulation applied to the earliest feature embedding stage, which biases low-level representations toward boundary cues. This involves a deterministic operator, where the modulated feature map is defined as F˜1 = F1 + λinj · Π(E), utilizing nearest-neighbour spatial rescaling of E.

  2. Boundary alignment loss: An auxiliary supervision term that penalizes mismatch between the predicted mask gradients and the RGB edge prior, which helps sharpen contours and reduce boundary bleeding.

Boundary Supervision Objectives

To ensure accurate delineation, a multi-scale boundary-informed objective is proposed, extending the base diffusion segmentation supervision with three complementary terms across output scales:

  1. Focal structure loss: This fuses a boundary-weighted focal cross-entropy with a weighted Intersection over Union (IoU) term, using a spatially adaptive weight map w(y) that emphasizes pixels in the vicinity of object boundaries.

  2. Ground-truth edge loss: This term directly supervises the gradient field of the prediction by penalizing the l1 distance between the Sobel response of the predicted probability map and the Sobel response of the binary ground-truth mask.

  3. Uncertainty-aware loss: This term uses an uncertainty map (ui) to create a spatially-varying penalty weight, encouraging the network to reduce spatial uncertainty globally and converge to more decisive, sharply-bounded predictions, particularly during early diffusion timesteps.

Multi-scale Aggregation

The final training objective is defined as Ltotal = Lms, which unifies region accuracy, boundary sharpness, uncertainty minimisation, and image-grounded contour alignment into a single objective. This is achieved by evaluating the four loss terms jointly across a set of spatial scales S = S = 1.0, 0.5, 0.25 with corresponding normalized weights ω = ω = 1.0, 0.25, 0.125. The scale-aggregated objective is defined in Eq (18) to ensure that the full-resolution supervision dominates while still providing multi-scale gradient signals that stabilize training at early diffusion timesteps.

Evaluation and Results

Experiments across the CAMO, COD10K, and NC4K datasets demonstrate consistent improvements in detection quality. The model consistently outperforms existing state-of-the-art methods across all evaluated metrics (Sm, Fwβ, Em, and MAE), delivering sharper delineation of thin structures and protrusions while also minimizing false positives. Specifically on CAMO, Bi-CamoDiffusion achieved a 2.4% Sm gain over the single-scale baseline, and on NC4K, it was the only method to surpass 0.90 in Sm and 0.95 in Em, achieving a MAE below 0.03 against the state-of-the-art methods including CamoDiffusion.

Key Contributions

The main contributions of this work are summarized as follows:

A structurally informed extension of CamoDiffusion to improve boundary delineation during diffusion-based mask prediction.

A multi-scale boundary-informed training objective that promotes contour aligned predictions and stabilizes structural consistency.

Evaluation across multiple COD benchmarks demonstrating improved boundary fidelity and detection reliability.

The final training objective is defined in Eq (19): Ltotal = Lms. The hyperparameters used for configuration include λinj = 0.075, λgt-edge = 0.01, λual = 0.01, and λrgb = 0.005 on CAMO and COD10K training subsets respectively. The model is optimized with AdamW at a learning rate of 5 × 10−5 over 150 epochs on two NVIDIA A100 GPUs of 24GB each. The full objective achieves the best results across all metrics, with Lrgb providing the final gain that consolidates all boundary supervision signals.

The gist: Bi-CamoDiffusion introduces a parameter-free prior injection process and a multi-scale boundary-informed loss function to enhance structural consistency and sharp contour delineation in camouflaged object detection.

Improvements for AI systems

Here are the specific improvements for AI systems based on Bi-CamoDiffusion, along with what those improved systems can achieve:

  1. Enhance object boundary delineation in visually ambiguous scenes (e.g., low contrast or complex textures).

  2. Reduce false positive detections by enforcing structural consistency between predicted masks and input image edges.

  3. Improve detection of thin structures and protrusions by explicitly incorporating early-stage edge priors into the diffusion process without altering the backbone architecture.

  4. Achieve more robust predictions in camouflaged environments under domain shifts (e.g., moving from training data to unseen benchmarks like NC4K).

  5. Generate more precise spatial support for segmentation masks, preventing boundary bleeding and fragmented object detections caused by weak visual transitions.

These improvements enable the system to perform high-fidelity segmentation in challenging camouflaged scenarios where traditional methods fail due to low foreground-background separability or structural ambiguity.

Sources

Related papers