C3Ghost Module in YOLO Detectors
- C3Ghost is a lightweight variant of the C3 block that leverages GhostNet-based ghost convolutions to efficiently generate feature maps.
- It replaces dense bottlenecks with cheap linear transformations like depthwise convolutions, reducing parameters and FLOPs in the process.
- Empirical evaluations show that while C3Ghost improves precision and speed, it may introduce trade-offs in localization and recall metrics.
Searching arXiv for papers on "C3Ghost module" and related YOLO architectures. C3Ghost is a lightweight convolutional module used in YOLO-like object detectors as a GhostNet-derived replacement for standard C3 or related bottleneck aggregation blocks. Across recent detector variants, it preserves the cross-stage partial macro-structure of the C3 family while replacing standard bottlenecks with GhostBottleneck or GhostConv-based operators so that feature maps are generated with fewer dense convolutions and more cheap linear transformations, typically depthwise operations; in reported uses, this is intended to reduce parameters and floating-point cost while preserving, or in some cases improving, feature representation quality (Ma et al., 2024, Joctum et al., 7 Jul 2025, Chen et al., 21 Jul 2025).
1. Definition and architectural identity
In the detection literature summarized here, C3Ghost is not a single immutable layer definition but a family of GhostNet-based C3 variants. The common pattern is stable: a CSP/C3-style block splits channels into two paths, processes one branch through one or more GhostBottleneck units instead of standard bottlenecks, concatenates the transformed and bypass branches, and fuses them with a final convolution. In Fostc3net, the authors state that “The original YOLOv5 Neck network's Bottleneck is substituted with the innovative C3Ghost module from GhostNet [11],” which directly identifies C3Ghost as a C3-like neck block with Ghost-based internal transforms (Ma et al., 2024).
YOLO-APD makes the same structural identification in a YOLOv8 setting. There, C3Ghost keeps the CSP macro-structure of the original C3/C2f family but replaces the standard bottlenecks inside the block with GhostBottleneck units, and the final fused output is produced by a CBM operator, i.e. Conv–BN–Mish (Joctum et al., 7 Jul 2025). The YOLOv9s-based steel-defect detector describes the same idea from a different baseline: the heavier RepNCSPELAN4 block is replaced by a C3Ghost module constructed from GhostConvBottleneck units, again with a shortcut path, concatenation, and final fusion convolution (Chen et al., 21 Jul 2025).
This convergence across implementations suggests that “C3Ghost” functions less as a proprietary block name than as a design pattern: a C3-style aggregation shell whose computational core is replaced by GhostNet-style cheap feature generation.
2. Internal mechanism and feature-generation principle
The key primitive inside C3Ghost is Ghost convolution. Rather than generating all output channels through a full dense convolution, a standard convolution first produces a smaller set of intrinsic feature maps, after which lightweight linear transformations generate additional “ghost” feature maps. In the YOLO-APD description, this is written as
followed by
and then
where is the input feature map, is a standard convolution kernel producing intrinsic channels, and denotes cheap linear transforms such as depth-wise convolutions (Joctum et al., 7 Jul 2025).
Fostc3net describes the same mechanism verbally: “Initially, a set of feature maps is generated through standard convolutional operations. Subsequently, these feature maps are subject to a cheap operation to produce redundant feature maps, generating additional Ghost feature maps. This step utilizes depthwise (DW) convolutions derived from simple linear transformations. Finally, the feature maps generated by the standard convolution are concatenated with those generated through the cheap operation” (Ma et al., 2024). The steel-defect detector gives the same operational interpretation, emphasizing that GhostConv generates only a subset of intrinsic features by expensive convolution and synthesizes the remaining channels via cheap operations (Chen et al., 21 Jul 2025).
Inside the broader C3Ghost shell, this GhostConv principle is embedded in a CSP dataflow. A representative conceptual flow, reconstructed explicitly in the cited detector papers, is: split the input channels, send one branch through a sequence of GhostBottleneck units, keep the second branch as a shortcut or lightweight convolutional path, concatenate the two branches, and apply a final fusion convolution (Ma et al., 2024, Joctum et al., 7 Jul 2025). The residual Add inside GhostBottleneck is preserved in the YOLO-APD formulation, and the CSP-style split–concat pathway is preserved at the block level (Joctum et al., 7 Jul 2025).
The exact micro-implementation is architecture-dependent. YOLO-APD explicitly couples C3Ghost with Mish activation and CBM layers (Joctum et al., 7 Jul 2025). Fostc3net does not provide an explicit layer-by-layer kernel listing for C3Ghost itself, although it reconstructs the module from known YOLOv5 C3 structure and GhostNet principles (Ma et al., 2024). The YOLOv9s paper similarly provides the main branch/shortcut/concat/fuse schematic rather than a full operator-by-operator derivation (Chen et al., 21 Jul 2025).
3. Placement in detection architectures
The position of C3Ghost within a detector varies substantially across model families. In Fostc3net, C3Ghost is a neck-only replacement. The backbone is modified separately with FasterNet blocks, while the neck’s original C3 modules adopt Ghost-based bottlenecks. The paper therefore assigns C3Ghost specifically to feature fusion, not to early feature extraction, and ties its role to reducing FLOPs in the feature channel fusion process (Ma et al., 2024).
YOLO-APD expands the scope of use. In that architecture, C3Ghost appears in both backbone and neck. In the backbone, it replaces ordinary C3/C2f bottleneck stacks where deep feature extraction occurs. In the IGD neck, it is used after multi-scale fusion operators such as SimFusion3in, SimFusion4in, and AdvPoolFusion, on both the descending and ascending paths. Thus, in YOLO-APD, C3Ghost is not merely a local neck optimization but a core structural element across the feature hierarchy (Joctum et al., 7 Jul 2025).
The YOLOv9s steel-defect detector applies C3Ghost as a replacement for RepNCSPELAN4 blocks in several locations of both backbone and neck while retaining the general YOLOv9s structure. The stated rationale is that these positions carry mid-level and high-level features with larger channel counts, so replacing heavy blocks there yields larger reductions in GFLOPs and parameters (Chen et al., 21 Jul 2025).
These placements reveal a common implementation logic. C3Ghost is typically deployed where channel widths are sufficiently large that channel redundancy is high and where dense convolutions would otherwise dominate inference cost. However, the papers also show that there is no single canonical placement policy: neck-only use, backbone-plus-neck use, and replacement of larger ELAN/CSP-style blocks all appear in practice.
4. Reported empirical behavior
The empirical literature represented here does not support a single universal performance profile; instead, it shows a recurring pattern of improved efficiency and precision-oriented behavior, with possible trade-offs in stricter localization metrics or recall unless compensated by other modules or losses.
| Model context | Placement of C3Ghost | Reported effect |
|---|---|---|
| Fostc3net | Neck | CSGhost “performs best in terms of accuracy”; full model reaches 5.212 Params (M), 12.88 GFLOPs, 0.988 mAP@0.5 |
| YOLO-APD | Backbone and neck | C3Ghost is described as computationally efficient locally, while full YOLO-APD reaches 77.7% [email protected]:0.95 and 100 FPS |
| YOLOv9s steel defect detector | Backbone and neck | C3Ghost alone changes 9.60M to 8.91M Params, 38.7 to 34.5 GFLOPS, and 94.3 to 131.5 FPS |
The clearest isolated ablation appears in the YOLOv9s steel-defect study. Relative to baseline YOLOv9s, adding C3Ghost alone changes Precision from 70.7 to 78.4, Recall from 77.3 to 70.3, [email protected] from 78.1 to 79.5, FPS from 94.3 to 131.5, GFLOPS from 38.7 to 34.5, and Params from 9.60M to 8.91M (Chen et al., 21 Jul 2025). This is a concrete example of the frequently reported precision–recall trade-off: the model becomes lighter and substantially faster, [email protected] improves, but recall decreases.
Fostc3net reports a whole-model comparison rather than a clean C3Ghost-only system. The full model, which integrates C3Ghost, FasterNet, and wIoU v3, changes from 7.102 Params (M), 14.82 GFLOPs, 0.987 [email protected], 0.714 [email protected]:0.95, 0.968 Precision, and 0.977 Recall in YOLOv5 to 5.212 Params (M), 12.88 GFLOPs, 0.988 [email protected], 0.711 [email protected]:0.95, 0.979 Precision, and 0.981 Recall in Fostc3net (Ma et al., 2024). The accompanying ablation states that the “CSGhost structure performs best in terms of accuracy,” and that “performance of CSGhost, based on neck modification, exceeds that of Fastnet on the backbone network” (Ma et al., 2024). At the same time, the paper states that adding Fastnet and CSghost improved precision but caused a decrease in the [email protected]–.95 metric, which was then mitigated by the wIoUv3 loss (Ma et al., 2024).
YOLO-APD presents a different but equally important lesson. It describes C3Ghost as a computationally efficient replacement designed to lower parameters and FLOPs locally, but the full model includes SimSPPF, SimAM, IGD, and Mish; as a result, whole-model cost increases relative to the YOLOv8 baseline even though detection accuracy improves. At the whole-model level, YOLO-APD versus YOLOv8 is reported as 77.7% versus 71.5% [email protected]:0.95, 97.0% versus 93.1% [email protected], 100 FPS versus approximately 120 FPS, 76.5G versus 67.7G FLOPs, and 24.16M versus 20.04M Params (Joctum et al., 7 Jul 2025). This demonstrates that the local efficiency of C3Ghost should not be conflated with guaranteed global model compression.
5. Interactions with surrounding modules
C3Ghost is rarely deployed in isolation. Its practical significance often emerges through interaction with other architectural or optimization components. In Fostc3net, the neck is lightweighted by C3Ghost while the backbone is lightweighted by FasterNet, which uses Partial Convolutions to process only a portion of the input channels. The loss is simultaneously changed to wIoU v3 in order to mitigate the decline of the [email protected]–.95 metric introduced by the structural lightweighting (Ma et al., 2024). In that design, C3Ghost reduces cost in feature fusion, FasterNet reduces cost in backbone extraction, and wIoU v3 compensates for localization-quality degradation.
YOLO-APD couples C3Ghost with SimAM, SimSPPF, IGD, and Mish. The paper presents these elements as synergistic: SimSPPF enhances multi-scale feature extraction, SimAM refines feature representations without additional parameters, the IGD-inspired neck improves feature fusion, Mish supports better optimization, and C3Ghost provides the lightweight local processor that makes repeated fusion operations more computationally manageable (Joctum et al., 7 Jul 2025). The role of C3Ghost inside the IGD neck is especially explicit: numerous multi-scale fusion points would be more expensive with standard C3 bottlenecks, so C3Ghost acts as the low-cost feature-processing block after fusion.
In the steel-defect detector, C3Ghost is combined with SCConv and CARAFE. SCConv reconstructs spatial and channel dimensions to reduce redundancy, while CARAFE performs content-aware upsampling for finer restoration of high-resolution defect regions. The paper’s ablation indicates that C3Ghost is the main driver of speed and parameter reduction, while SCConv and CARAFE recover part of the recall loss and further refine detection performance (Chen et al., 21 Jul 2025).
A plausible implication is that C3Ghost is most effective not as a standalone universal remedy but as a lightweight feature-processing primitive embedded in a broader architecture-specific compensation strategy. The compensatory element may be an improved loss, an attention operator, a different fusion mechanism, or an upsampling module.
6. Variability, limitations, and common misconceptions
One common misconception is that C3Ghost denotes a fully standardized module. The cited detector papers do not support that interpretation. Fostc3net reconstructs the block from YOLOv5 C3 and GhostNet concepts and explicitly notes that the paper does not give a line-by-line layer listing for C3Ghost itself (Ma et al., 2024). YOLO-APD specifies a C3Ghost tied to CBM and Mish (Joctum et al., 7 Jul 2025). The YOLOv9s defect detector presents it as a replacement for RepNCSPELAN4 using GhostConvBottleneck blocks (Chen et al., 21 Jul 2025). The shared identity is therefore structural rather than implementation-identical.
A second misconception is that C3Ghost always improves every relevant metric. The reported results are more nuanced. In Fostc3net, the structural lightweighting that includes C3Ghost improves precision but causes a decline in [email protected]–.95 until wIoUv3 is introduced (Ma et al., 2024). In the steel-defect detector, C3Ghost alone raises Precision and [email protected] but lowers Recall (Chen et al., 21 Jul 2025). These observations indicate that the module’s efficiency gain may alter the detector’s precision–recall balance and localization behavior.
A third misconception is that C3Ghost necessarily reduces whole-model parameters and FLOPs. The local computational logic of Ghost convolution does point in that direction, but whole-architecture outcomes depend on what else is added. YOLO-APD is the clearest counterexample: C3Ghost is described as computationally efficient, yet the full model has higher FLOPs and parameter count than baseline YOLOv8 because the additional SimSPPF, IGD, and other components more than offset the local savings (Joctum et al., 7 Jul 2025).
Finally, the term “ghost” in C3Ghost is unrelated to the use of “ghost” in theoretical physics. In multimetric gravity, “ghost-free” refers to the absence of unwanted scalar degrees of freedom such as the Boulware–Deser ghost (Nomura et al., 2012). In higher-derivative mechanics, “ghost-free” concerns the elimination of Ostrogradsky instabilities by a hierarchy of degeneracy conditions (Motohashi et al., 2017). Those usages are terminologically unrelated to the GhostNet-derived feature-generation strategy used in C3Ghost.
7. Significance in lightweight detection
Within lightweight object detection, C3Ghost occupies a specific design niche. It preserves the familiar C3/CSP pattern used in YOLO-family detectors while replacing standard dense bottlenecks with Ghost-based operators that exploit feature redundancy. This makes it attractive when architectural continuity matters: it can be integrated into existing C3, C2f, CSP, ELAN, or RepNCSPELAN-like regions without redesigning the entire detector (Ma et al., 2024, Chen et al., 21 Jul 2025).
The papers surveyed here consistently frame C3Ghost as a module for resource-constrained or real-time deployment. Fostc3net targets mobile devices for transmission-line monitoring (Ma et al., 2024). YOLO-APD emphasizes real-time pedestrian detection on geometrically complex roads, reporting 100 FPS while using C3Ghost as part of its efficiency strategy (Joctum et al., 7 Jul 2025). The steel-defect detector presents C3Ghost as a means of increasing inference speed while retaining or slightly improving mAP on multi-scale industrial defects (Chen et al., 21 Jul 2025).
Taken together, these reports establish C3Ghost as a recurrent architectural motif in contemporary lightweight detection: a GhostNet-based reparameterization of C3-style processing blocks that usually lowers per-block computational burden, often improves precision-oriented metrics, and typically requires complementary architectural or optimization choices when strict localization quality or recall must also be preserved.