Project Status: Complete | Dataset: AgriPest (Tiny-Object) | Key Finding: Hybrid v10v11 achieves near-YOLOv8 accuracy with 32% fewer parameters.

This independent study investigates the structural modularity and scaling behavior of modern YOLO architectures. By treating YOLO versions (v8, v10, v11) not as fixed models, but as modular systems, we created hybrid architectures (e.g., v10 Backbone + v11 Head) to identify novel performance and efficiency trade-offs on a challenging, tiny-object dataset.

The core hypothesis tested is whether architectural efficiency modules (from v10/v11) can be successfully combined with robust feature extraction mechanisms (from v8) to create a new Pareto-optimal model.

This experiment explores the influence of different Backbone–Head pairings on: Accuracy (mAP50-95), Compute Cost (FLOPs, Params), and Real-World Latency ($\text{Infer}(\text{ms})$).

All models were trained on the AgriPest dataset (13 pest classes) to focus the analysis on tiny-object detection performance. Inference benchmarking was conducted on a consistent Intel i5-12500 CPU setup to capture real-world latency, rather than theoretical GPU speed.

The table below summarizes the key validation metrics, where the best result for a given metric is bolded.

The v10v11 hybrid (YOLOv10 Backbone + YOLOv11 Head) is the most significant finding. It achieved a $\text{mAP50-95}$ of 0.409, which is statistically indistinguishable from the leading YOLOv8 baseline ($\text{mAP50-95}=0.416$), yet it utilizes only 2.06 million parameters—a 32% reduction compared to YOLOv8's 3.01M. This confirms the efficacy of hybridization in finding a superior theoretical efficiency point.

A major counter-intuitive result is the lack of correlation between theoretical complexity (FLOPs/Params) and practical CPU inference latency:

• YOLOv11 has the lowest FLOPs (6.5G) but is only marginally faster than the most parameter-heavy model ($\text{YOLOv8} \rightarrow 13.9\text{ms}$).
• The v10v11 hybrid, with the lowest parameter count (2.06M), is the slowest model at $27.9\text{ms/img}$.

This is visually confirmed by the Metric Correlation Heatmap, which shows a high negative correlation ($\approx -0.89$) between Parameters and Inference Time. This suggests that on CPU hardware, architectural overhead, memory access patterns, and kernel optimization dominate the runtime cost over raw computation count.

Models leveraging the robust, dense-feature YOLOv8 Head (YOLOv8, v11v8) consistently delivered strong $\text{mAP50-95}$ performance, suggesting that for high-resolution, fine-grained tasks like tiny-object detection, the complexity of the feature aggregation head is more critical than the efficiency of the backbone.

These plots illustrate the classic trade-off space, showing how the models cluster. The ideal model would sit in the top-right (high accuracy) and bottom-left (low compute/latency) quadrants.

Larger area indicates higher overall detection performance. YOLOv8 and the v10v11 hybrid dominate the top quadrant.

Lower values are better for efficiency. Note how the latency ($\text{Infer}(\text{ms})$) for the v10v11 hybrid drastically pulls its overall efficiency profile outwards, despite having the best parameters (Params(M)).

The visual proof of the efficiency paradox: a strong negative correlation between Parameters and Inference Time.

1. In-Depth Profiling: Conduct detailed profiling (e.g., PyTorch Profiler, TensorBoard) specifically to diagnose the architectural causes of the $\text{Params}-\text{Inference}$ discrepancy observed on CPU hardware.
2. Pareto Frontier Analysis: Plot the $\text{mAP50-95}$ vs. $\text{FLOPs}$ to formally identify which models occupy the Pareto Frontier for this tiny-object task.
3. Cross-Dataset Validation: Replicate the experiment on a broader dataset (e.g., a COCO subset) to validate if the hybridization findings are generalizable beyond the AgriPest domain.
4. Hardware Benchmarking: Benchmark the top hybrids on different hardware (e.g., GPU/Edge TPU) to see if the latency bottleneck is CPU-specific.

• Hybridization is Valid: Recombining components across YOLO generations is a viable architectural research methodology.
• The v10v11 Hybrid provides a new, highly parameter-efficient architecture approaching state-of-the-art accuracy on tiny objects.
• Theoretical vs. Real-World Efficiency: Raw FLOPs/Parameter counts are insufficient predictors of latency on modern CPU hardware; deep architectural factors must be considered.

Author: Y. R. A. V. R — Hyderabad, India 🔗 LinkedIn | GitHub

• Ultralytics YOLOv8 → v11 Releases
• SCDown and C3k2 modules (YOLOv10 paper)
• CIB Blocks and Hybrid Heads (YOLOv11)
• AgriPest Dataset