YOLO-IOD: Specialized YOLO Detectors

• YOLO-IOD is a suite of three distinct object detection frameworks derived from YOLO, each tailored for continual, IoT, or oriented detection challenges.
• It employs specialized techniques—such as CPR, PhantomConv, and DirIoU—to enhance incremental learning, low-light multimodal processing, and orientation precision.
• Each variant demonstrates strong real-time performance with optimized efficiency, making them ideal for application-specific detection tasks.

YOLO-IOD denotes three distinct, technically rigorous object detection frameworks based on the YOLO family, each addressing a different challenge in the domain: (1) real-time incremental object detection for continual learning, (2) compact, IoT-ready obscured detection using multimodal data, and (3) efficient uniform-size oriented object detection with directional awareness. These systems are unified by their derivation from YOLO, their focus on real-time applications, and their significant architectural and methodological divergences from standard YOLO for specialized use cases.

1. Problem Definitions and Framework Overview

The term YOLO-IOD has been applied to three non-overlapping yet related frameworks:

• Incremental Object Detection YOLO-IOD (Zhang et al., 28 Dec 2025): A real-time framework for incremental object detection that preserves and extends performance as new classes are introduced in sequential training phases, minimizing catastrophic forgetting. Built atop YOLO-World, it introduces principled mechanisms for pseudo-label management, parameter isolation, and knowledge distillation.
• Obscured Detection YOLO-IOD (Mukherjee et al., 2024): An ultra-compact YOLO Phantom model, designed for embedded and IoT settings, which incorporates the novel Phantom Convolution block for parameter-efficiency and robust performance in low-light, multimodal (RGB+infrared) scenarios.
• Uniform Oriented Detection YOLO-IOD ("YUDO") (Nedeljković, 2023): A minimal adaptation of YOLOv7-tiny dedicated to detecting directed, fixed-size objects by regressing position and orientation (angle), leveraging a novel DirIoU metric sensitive to orientation.

Each framework features substantial modifications to YOLO’s canonical detection pipeline, customized annotation formats, unique losses or matching criteria, and application-specific deployment considerations.

2. YOLO-IOD for Incremental Object Detection (Zhang et al., 28 Dec 2025)

This YOLO-IOD instance addresses catastrophic forgetting in continual learning by integrating three orthogonal mechanisms, evaluated on both conventional and the stricter LoCo COCO benchmarks.

2.1 Knowledge Conflict Taxonomy

Three interdependent sources of forgetting are formally identified:

• Foreground–Background Confusion: Unlabeled instances of old or future classes in current-phase datasets are treated as background, especially exacerbated by heavy augmentations.
• Parameter Interference: Shared convolutional parameters, if indiscriminately updated, overwrite features critical for previously learned classes.
• Misaligned Knowledge Distillation: Dissonance between teacher and student detectors across incremental phases causes improper knowledge transfer, since their output spaces partially mismatch.

The framework introduces precise quantifications for these conflicts, e.g., Foreground–Background Confusion is measured as the expected rate at which prior/future objects are suppressed as background.

2.2 Architecture and Optimization

• Base Model: YOLO-World, combining a visual encoder $f_v$ and text encoder $f_t$ fused via RepVL-PAN.
• Stage-wise, Parameter-Efficient Fine-Tuning (PEFT): Only a small, systematically selected fraction (12–20%) of kernels is adapted at each stage (see below).
• Classifier Expansion: The detection head’s class prototypes are dynamically expanded to incorporate new categories without disrupting representations for existing ones.

2.3 Core Modules

• Conflict-Aware Pseudo-Label Refinement (CPR): Weighted pseudo-label loss combining focal-style confidence weighting and entropy regularization. Additionally, unannotated detections are assigned to clusters in a generalized open vocabulary, forming pseudo-supervision for unknown classes via frequency-weighted k-means.
• Importance-Based Kernel Selection (IKS): Employs Fisher Information to measure the current and accumulated importance of each convolutional kernel:

$$\Delta I_t(w^k) = I_t(w^k) - \rho \sum_{i=1}^{t-1} I_i(w^k)$$

Selecting only the top-K kernels with highest $\Delta I_t$ for fine-tuning preserves prior-task representations.

• Cross-Stage Asymmetric Knowledge Distillation (CAKD): Enacts dual-teacher asymmetric distillation. The student’s features are processed by both the previous and current heads, and the KD objective symmetrizes losses between old and new class outputs using a focal-weighted composite of classification and regression losses.

2.4 Benchmark and Results

A new LoCo COCO protocol is established to prevent image overlap between incremental phases, avoiding data leakage. YOLO-IOD achieves:

• Single-step (40+40 COCO): AP = 53.0, Absolute Gap = 1.5, Relative Gap = 2.7%
• Multi-step (20-20): Relative Gap reduced to 5.1%, outperforming RGR and ERD with identical YOLO-World backbones

It robustly closes the forgetting gap versus earlier YOLO-based IOD systems, while retaining real-time speed. All modules are ablated for independent contribution; LoCo COCO reveals that previous approaches significantly overestimated performance due to split leakage.

3. YOLO-IOD as Efficient IoT/Obscured Detection (Mukherjee et al., 2024)

This YOLO-IOD refers to “YOLO Phantom”—a compact, multimodal detector.

3.1 Model Architecture

• Backbone: Variant of YOLOv8n, integrating four-stage CSP C2fi blocks with shortcut and filter reduction, and a SPPF module. Phantom Convolution modules are inserted to replace heavier standard convolutions.
• Detection Head: Three decoupled scale-specific heads, final layer using PhantomConv for parameter efficiency.
• Phantom Convolution Block: Two-step process.
  • Group Convolution for channel split efficiency.
  • Depthwise-Separable Convolution for per-channel locality.
• Total Network Size: $\approx$1.82M parameters (vs. 3.20M YOLOv8n), $\approx$6.07 GFLOPs.

3.2 Multimodal and Training Protocol

• Input: Channel-alternating (“fusion by joint training”) of RGB and gray-scale IR frames, obviating the need for dual-stream networks.
• Loss: Standard YOLO detection loss; no explicit cross-modality regularization.
• Training: Pretrained on COCO, gradual unfreezing, cosine LR schedule.

3.3 Efficiency and Deployment

• Accuracy: [email protected]: 19.72% (RGB), 24.82% (IR) – on par with original YOLOv8n
• FPS on RPi 4B: 15.5 (RGB), 16.6 (IR) – a 14–17% improvement
• System Integration: Exported to ONNX and deployed with ncnn/OpenVINO engines, integrated with AWS for event notification and containing hardware/thermal fail-safes.

This instantiation demonstrates near-baseline accuracy under both thermal and visual-only regimes, significant parameter and floating-point operation reduction, and seamless field deployment.

4. YOLO-IOD for Uniform Directed Object Detection (YUDO, (Nedeljković, 2023))

YUDO represents a drastic but minimal reworking of YOLOv7-tiny for fixed-size, oriented target detection.

4.1 Architectural Modification

• Detection Head: Collapse from three to a single 16×16 head (one per feature grid cell), anchor-free, none of the usual (width, height) regression channels; output vector is $(x, y, \theta, obj, cls_{abdomen}, cls_{bee})$.
• Input: 512×512 images, one object per cell, matching the fixed spatial frequency of the observed objects.

4.2 Annotation and Output Format

• $(x, y)$: normalized cell offsets for centroid (sigmoid)
• $\theta \in [0, 2\pi)$: absolute direction angle (ReLU + modulo)
• No $(w, h)$ outputs or regression; all objects use a fixed box size.
• Class labels: 'abdomen' (circle; no orientation) and 'bee' (with orientation).

4.3 DirIoU: Oriented Matching and NMS

• SkewIoU: Standard area IoU for rotated boxes $P$, $Q$.
• Direction Correction Factor: $$\mathrm{DirCorr}(\Delta \theta) = (1 + \cos \Delta \theta)/2$$
• DirIoU: $$\mathrm{DirIoU}(P, Q) = \mathrm{SkewIoU}(P, Q) \cdot \mathrm{DirCorr}(\Delta \theta)$$

This formulation ensures that overlapping detections with reversed angles are considered fully separable (DirIoU=0 when $\Delta \theta = \pi$), influencing both matching ($\geq 0.3$ for mAP) and NMS.

4.4 Training and Performance

• Dataset: 13,908 training, 1,392 validation images from the Honeybee Segmentation and Tracking Dataset.
• Losses: $L_{xy}$ (MSE), $L_\theta$ ($1 - \cos(\hat\theta - \theta)$), standard binary cross-entropy for objectness and class.
• Final Results: mAP@30 (DirIoU $\geq 0.3$): bee = 85.1%, abdomen = 55.9%, overall = 70.5% using $\sim$6M parameters, 13.1 GFLOPS, $>$50 FPS on RTX2070 Super.

This “editor’s term”: YUDO demonstrates that, given uniform object sizes and consistent angular semantics, YOLO can be dramatically simplified and made orientation aware with minimal cost.

5. Comparative Performance and Domain-Specific Implications

YOLO-IOD Type	Core Technique	Parameter Count	Main Use Case	SOTA/Benchmarks
Incremental (2512...)	CPR, IKS, CAKD	YOLO-World(x)	Real-time continual/incremental detection	SOTA LoCo COCO, $<$3% RelGap
Obscured/IoT (2402...)	PhantomConv, multimodal train	1.82M	Edge/IoT multimodal, low-light detection	mAP~20–25%, +14–17% FPS vs. base
Oriented/YUDO (2308...)	DirIoU, anchorless head	$\sim$6M	Uniform size, orientation-sensitive objects	mAP@30=70.5%, $>$50 FPS RTX2070

YOLO-IOD, as instantiated in these domains, enables real-time learning and inference while addressing continual learning, hardware efficiency, and orientation precision. Notably, technical innovations such as DirIoU and CAKD have architectural and algorithmic implications for broader object detection research.

6. Limitations and Future Directions

• Incremental YOLO-IOD: Dependence on open-vocabulary teacher for CPR; IKS kernel ratio requires tuning; dual teachers in CAKD increase memory overhead.
• YOLO Phantom: No explicit cross-modal loss; “fusion” occurs by joint training, which may limit theoretical fusion capacity.
• YUDO: Assumes fixed-size, minimally-overlapping objects; direct transfer to highly variable or dense scenes may degrade performance.

Suggested future work includes adaptive, meta-learned kernel selection (IKS tuning), dynamic unknown cluster discovery in CPR, generalization of DirIoU for variable-size objects, and plug-and-play dual-enhancement modules.

---

For technical implementation details and further results, refer to the source papers: (Zhang et al., 28 Dec 2025, Mukherjee et al., 2024), and (Nedeljković, 2023).