FBRT-YOLO: Faster and Better for Real-Time Aerial Image Detection

This paper addresses a practical bottleneck in real-time aerial image detection: how to improve small-target detection accuracy without sacrificing inference speed and computational efficiency on resource-constrained embedded flight devices. The authors frame the core research question as balancing (i) the difficulty of detecting very small objects in high-resolution aerial imagery and (ii) the need for fast inference under limited compute budgets. This matters because many standard real-time detectors (especially those optimized for low-resolution natural images) either lose small-object information during downsampling/feature extraction or become too computationally expensive when adapted to high-resolution aerial inputs.

The proposed solution, FBRT-YOLO, is a family of real-time detectors built around two lightweight modules designed specifically to improve small-object perception while keeping the model efficient. The Feature Complementary Mapping Module (FCM) targets an information mismatch problem: shallow layers contain precise spatial location cues, while deeper layers contain richer semantic cues, but conventional backbones do not integrate shallow spatial information into deeper representations effectively. The Multi-Kernel Perception Unit (MKP) targets another issue: small objects may occupy only a few pixels and can be “washed out” by receptive-field limitations and downsampling; additionally, small objects exist at multiple scales and are often surrounded by complex backgrounds. MKP is intended to enhance multi-scale feature perception with minimal structural overhead.

Methodologically, the paper follows a standard deep learning object detection evaluation pipeline. FBRT-YOLO is implemented as a YOLO-family detector with modifications to the backbone/neck design. Training is performed for 300 epochs using SGD with momentum 0.937, weight decay 0.0005, batch size 4, and initial learning rate 0.01. Experiments are run on three aerial detection benchmarks: VisDrone (2018), UAVDT (2018), and AI-TOD (2021). Hardware details are provided: training uses an NVIDIA GeForce RTX 4090 GPU, while inference speed is measured on a single RTX 3080 GPU. The paper reports detection quality using average precision (AP) and AP50/AP75, and efficiency using GFLOPs, parameter counts, and FPS.

The key architectural contributions are described as follows. FCM is embedded into each stage of the backbone. It uses a channel split with ratio $\alpha$\u0007 to divide input feature channels into two branches: one branch is processed with a standard $3\times 3$ convolution to extract richer semantic/channel information, while the other branch uses point-wise convolution to preserve more shallow spatial positional information. The module then performs complementary mapping via channel interaction (depthwise convolution followed by global average pooling and sigmoid to generate channel weights) and spatial interaction (a convolution + BN + sigmoid to generate spatial attention weights). Finally, it aggregates the two branches using element-wise weighted fusion: $$X_{\text{FCM}}=(X_C\otimes {\omega }_2)\oplus (X_S\otimes {\omega }_1).$$ FCM’s intent is to transfer spatial location information into deeper semantic representations, improving alignment and localization of small targets.

MKP is introduced in the final (fourth) backbone stage. It replaces the final downsampling layer and reduces the corresponding detection heads, aiming to simplify the network while improving multi-scale perception. MKP concatenates depthwise convolutions with different kernel sizes (the paper’s experiments explore kernel-size configurations) and uses point-wise convolutions between scales to integrate cross-scale spatial relationships. The paper summarizes MKP as: $$X^{\prime }=T_{2k+1}(A(\cdots A(T_k(X))\cdots )).$$ In practice, the authors set $k=3$ (implying multi-kernel sizes such as $3,5,7$ in their ablations).

The paper’s main quantitative results are reported on VisDrone in Table 1, comparing FBRT-YOLO variants (N/S/M/L/X) against multiple real-time detectors (YOLOv5-L, YOLOv8-N/S/M/L/X, YOLOv9-M, YOLOv10-S/L/X, and RT-DETR-R34/R50). Across model sizes, FBRT-YOLO improves the accuracy–efficiency trade-off. For example, on VisDrone: - FBRT-YOLO-N achieves AP $20.2\%$ and AP50 $34.4\%$ with 0.9M parameters and 6.7 GFLOPs, running at 192 FPS. - FBRT-YOLO-S achieves AP $25.9\%$, AP50 $42.4\%$ with 2.9M parameters and 22.9 GFLOPs at 143 FPS. - FBRT-YOLO-M achieves AP $28.4\%$, AP50 $45.9\%$ with 7.2M parameters and 58.7 GFLOPs at 94 FPS. - FBRT-YOLO-L achieves AP $29.7\%$, AP50 $47.7\%$ with 14.6M parameters and 119.2 GFLOPs at 70 FPS. - FBRT-YOLO-X achieves AP $30.1\%$, AP50 $48.4\%$ with 22.8M parameters and 185.8 GFLOPs at 52 FPS.

The authors also provide relative improvements in terms of parameter/GFLOPs reductions and AP gains versus YOLOv8 and YOLOv10 baselines. For instance, they state that FBRT-YOLO-N/S reduces parameter counts by 72% and 74% compared to YOLOv8-N/S while improving AP by 0.6% and 2.3%. For medium models, FBRT-YOLO-M reduces GFLOPs by 26% (vs YOLOv8-M) and 23% (vs YOLOv9-M) while improving AP by 1.3% and 1.2%. For large models, FBRT-YOLO-X is reported to have 66% and 23% fewer parameters than YOLOv8-X and YOLOv10-X, with AP improvements of 1.2% and 1.4%.

On VisDrone, Table 2 compares against additional state-of-the-art methods, showing FBRT-YOLO reaching AP $30.1\%$ and AP50 $48.4\%$ and AP75 $31.7\%$, outperforming YOLOv3-SPP3 (AP $26.4\%$), DMNet (AP $29.4\%$), QueryDet (AP $28.3\%$), and CEASC (AP $28.7\%$).

On UAVDT (Table 3), FBRT-YOLO reports AP $18.4\%$, AP50 $31.1\%$, and AP75 $18.9\%$, exceeding CEASC (AP $17.1\%$, AP50 $30.9\%$) and other listed methods such as GLSAN (AP $17.0\%$) and GFL (AP $16.9\%$).

On AI-TOD (Table 4), the paper compares FBRT-YOLO-S against YOLOv8-S: FBRT-YOLO-S improves AP from $19.1\%$ to $20.2\%$ and AP50 from $43.6\%$ to $45.8\%$, while reducing parameters from 11.2M to 2.9M and GFLOPs from 28.6G to 22.9G. FPS increases slightly from 131 to 142.

The ablation study (Table 5) on VisDrone using YOLOv8-S as baseline evaluates the contributions of FCM, MKP, and a redundancy-reduction (RR) design. The baseline (no FCM, no MKP, no RR) has AP $23.6\%$, AP50 $39.6\%$, 11.2M parameters, and 28.6 GFLOPs. Adding RR alone yields AP $23.4\%$ (slightly lower) but reduces parameters to 9.1M and FLOPs to 25.5G. Adding FCM (with RR) yields AP $24.3\%$ and AP50 $40.6\%$ with 7.2M parameters and 23.2G FLOPs. Adding both FCM and MKP (and RR) yields the best ablation result: AP $25.9\%$, AP50 $42.4\%$ with 2.9M parameters and 22.9G FLOPs.

Additional ablations explore FCM’s split ratio $\alpha$ (Table 7), showing that configurations retaining more spatial information in deeper stages (e.g., $\alpha$ patterns like $(0.75,0.75,0.25,0.25)$) achieve higher AP50 and AP. Kernel-size choices for MKP (Table 8) indicate that too-small kernels limit receptive field, while too-large kernels introduce background noise; the mixed kernel configuration $(3,5,7)$ achieves AP $25.9\%$ and AP50 $42.4\%$ with 2.90M parameters and 22.9G FLOPs.

Limitations are not deeply quantified in the provided text. However, several apparent limitations follow from the methodology and reporting style: (1) the paper focuses on three benchmarks and does not report cross-dataset generalization beyond these; (2) inference speed is measured on a specific GPU setup, so real embedded deployment may differ; (3) the ablation and comparisons emphasize AP/AP50/AP75 and efficiency, but do not provide detailed statistical significance testing (e.g., confidence intervals) or robustness analyses across weather/illumination beyond qualitative heatmaps/visualizations; and (4) the paper does not clearly specify whether results are averaged over multiple runs with different seeds.

Practically, the results suggest that FBRT-YOLO is a strong candidate for real-time aerial perception pipelines where small-object detection is critical (e.g., drone-based monitoring, traffic/incident detection, and search-and-rescue). Who should care includes practitioners building on-device or near-device detection systems, researchers working on small-object detection in high-resolution imagery, and teams needing a YOLO-like architecture that improves the accuracy–efficiency trade-off without resorting to expensive multi-scale input pyramids.

Overall, the paper’s core contribution is a YOLO-family detector that combines spatial-semantic complementary mapping (FCM) with multi-kernel multi-scale perception (MKP), achieving improved AP while maintaining real-time performance across multiple model sizes on standard aerial benchmarks.