Is Bigger Always Better? Efficiency Analysis in Resource-Constrained Small Object Detection

Gabriel Kasmi, Kwame Mbobda-Kuate

In the world of deep learning, the prevailing wisdom is often "scale or fail." We’ve been conditioned by scaling laws to believe that larger models trained on more data will inevitably push the needle of performance. But does this "bigger is better" mantra hold true when you’re looking for tiny solar panels in drone imagery over Madagascar?

A new paper, Is Bigger Always Better? Efficiency Analysis in Resource-Constrained Small Object Detection, challenges this fundamental assumption. The authors demonstrate that in the data-scarce, resource-constrained world of Earth Observation (EO), the smallest models aren't just more efficient—they are actually more accurate.

TL;DR

In data-scarce Earth Observation tasks, standard scaling laws flip: the smallest model (YOLO11N) outperformed the largest (YOLO11X) in both absolute accuracy and efficiency. For detecting small objects like rooftop solar panels, prioritizing high input resolution over model size or dataset volume is the winning strategy.

What Problem Does This Paper Address?

Most computer vision research happens in "data-rich" environments like COCO or ImageNet. In these settings, scaling up model parameters consistently yields better results. However, Earth Observation (EO) is different. Annotating satellite or drone imagery requires domain expertise and is incredibly expensive, leading to structurally "data-scarce" datasets.

Furthermore, these models are often intended for deployment in developing countries or on edge devices where compute power is a luxury. The authors tackle a critical question for practitioners: Given a fixed resource budget, how should you allocate it between model complexity, dataset size, and image resolution?

Methods

The researchers conducted a systematic efficiency analysis using the YOLO11 family (from Nano to Extra-Large). They explored three scaling dimensions:

  1. Model Size: Five variants ranging from 2.6M to 57M parameters.
  2. Dataset Fraction: Training on 10%, 25%, 50%, and 100% of the available data.
  3. Input Resolution: 416px, 640px, and 1280px.

They defined a specific metric for Model Efficiency: $10 \times \frac{mAP_{50}}{ModelSize (MB)}$. This allows them to measure "performance per unit of storage," which is vital for edge deployment. They also analyzed the Pareto frontier in the joint accuracy–throughput (FPS) space to see if there was a genuine tradeoff between speed and precision.

Datasets

The study uses the OpenStat Madagascar rooftop PV dataset.

  • Nature: A crowdsourced collection of drone and satellite imagery.
  • Task: Detecting rooftop solar panels and boilers.
  • Challenge: These are "small objects." 64% of the annotated bounding boxes occupy less than 1% of the total image area.
  • Scale: The final filtered dataset contained 8,977 drone images.

Key Insights and Ideas

  • The Efficiency Inversion: In a shocking reversal of standard scaling laws, the smallest model (YOLO11N) achieved the highest absolute accuracy (mAP). It wasn't just "good for its size"; it was the best, period.
  • Resolution is King: Increasing input resolution was the most effective way to boost performance. Moving from 416px to 1280px provided a +120% efficiency gain. At low resolutions, adding more training data yielded almost no improvement because the objects were simply too blurry for the model to "see."
  • The Overparameterization Penalty: Large models (like YOLO11X) suffered in this data-scarce regime. With a high "overparameterization ratio" (params vs. training samples), these giants tended to memorize noise rather than generalize features, leading to worse performance than their smaller siblings.
  • No Tradeoff to Resolve: Usually, engineers expect to sacrifice accuracy for speed. Here, the small, high-resolution configurations were Pareto-dominant. They were simultaneously the most accurate and the fastest at inference, leaving no reason to choose a larger model.

Key Results

  • YOLO11N vs. YOLO11X: The Nano model was 24× more efficient than the Extra-Large model.
  • Top Performance: YOLO11N at 1280px resolution achieved the highest absolute $mAP_{50}$ of 0.617.
  • Data Hunger: Large models showed "pronounced data hunger." YOLO11X required the full dataset just to approach the performance that YOLO11N achieved with only 10% of the data.
  • Zero-Shot Failure: Models pre-trained on COCO had near-zero performance on this task (mAP < 0.004) before fine-tuning, proving that general-purpose computer vision features don't easily transfer to the unique top-down perspective of EO.

Limitations and Future Work

The authors acknowledge several bounds to their study:

  • Architecture Scope: The findings are specific to the YOLO11 family; Transformers or two-stage detectors might behave differently.
  • Geographic Bias: The data is limited to Madagascar. Results might vary in different urban environments.
  • Hardware Constraints: The maximum resolution (1280px) was limited by the 16GB VRAM of the NVIDIA A2 GPUs used. Higher resolutions might reveal even more dramatic gains.
  • Future Directions: The authors suggest exploring domain-adaptive pre-training (training on unlabeled satellite imagery first) to see if it can shift the "crossover point" where larger models become useful.

The takeaway for the EO community is clear: If you're working with limited data and small objects, don't waste your budget on massive models. Buy a better camera (or use higher resolution crops) and stick with the "Nano" variants. In this field, smaller isn't just leaner—it's smarter.

Summary updated

Abstract

Scaling laws assume larger models trained on more data consistently outperform smaller ones -- an assumption that drives model selection in computer vision but remains untested in resource-constrained Earth observation (EO). We conduct a systematic efficiency analysis across three scaling dimensions: model size, dataset size, and input resolution, on rooftop PV detection in Madagascar. Optimizing for model efficiency (mAP per unit of model size), we find a consistent efficiency inversion: YOLO11N achieves both the highest efficiency ( higher than YOLO11X) and the highest absolute mAP (0.617). Resolution is the dominant resource allocation lever ( 120% efficiency gain), while additional data yields negligible returns at low resolution. These findings are robust to the deployment objective: small high-resolution configurations are Pareto-dominant across all 44 setups in the joint accuracy-throughput space, leaving no tradeoff to resolve. In data-scarce EO, bigger is not just unnecessary: it can be worse.
BibTeX
@misc{kasmi2026is,
  author = {Gabriel Kasmi and Kwame Mbobda-Kuate},
  title = {{Is Bigger Always Better? Efficiency Analysis in Resource-Constrained Small Object Detection}},
  year = {2026},
  eprint = {2603.02142},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2603.02142},
}

Loading related content...

Figure