Bee Detection and Tracking at Hive Entrance using YOLO11 and ByteTrack

arXiv:2608.23213 · cs.CV · Submitted 2026-08-24 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: I'm Tom, and with me are Jane, Lu, senior AI researcher at Tsinghua, Meng, lead engineer at a mysterious AI startup and Lalam, the in-house Large Language Model.

Jane: Today's paper: "Bee Detection and Tracking at Hive Entrance using YOLO11 and ByteTrack".

Tom: This work presents an automatic bee entrance monitoring system based on YOLO11 transfer learning and ByteTrack tracking, investigating how data augmentation, backbone freezing,

Jane: First, who's behind it and why it matters.

Title and authors: Tom: So we’re starting with the title and authors of "Bee Detection and Tracking at Hive Entrance using YOLO11 and ByteTrack," which tells us exactly what tools they are combining to solve this monitoring problem. This paper is essentially showing how a specific combination of detection technology, YOLOv11, paired with a tracking method, ByteTrack, performs in a real-world scenario.

Jane: It’s important to understand that the authors are testing how these components interact under the stress of detecting small objects that move quickly and get blurry near an entrance. It’s about making sure the detection part is sharp enough to catch them, and the tracking part keeps them together even when they jump around.

Lu: The authors chose YOLOv11 because they noted its improved localization performance, which is exactly what you need when dealing with tiny, fast-moving targets; it seems like a logical choice for this kind of fine-grained vision task.

Meng: I wonder about the initial setup; they started both models with COCO pre-trained weights and then used transfer learning on their bee dataset to accelerate convergence, which makes sense if we don't have millions of bee images already.

Lalam: It seems like the authors are setting up a very targeted experiment by focusing specifically on these three variables—augmentation, freezing, and tracking parameters—to isolate which part of the system contributes most to the final counting accuracy.

The paper's summary: Tom: The summary of "Bee Detection and Tracking at Hive Entrance using YOLO11 and ByteTrack" boils down to them investigating how changing data augmentation, backbone freezing, and tracker parameters affects the detection and counting success rates for small bees. They are essentially trying to find the best recipe for a robust automatic monitoring system.

Jane: Simply put, they looked at how different ways of boosting the training data influence performance versus how different ways of structuring the neural network during training affects stability and accuracy in catching those bees. It’s a deep dive into optimizing every piece of the pipeline to handle real-world video noise.

Lu: The core finding they highlighted is that light augmentation performed better than heavy augmentation, which suggests that overwhelming the model with excessive data distortion actually hurts its ability to recognize those small bee features.

Meng: That result is interesting because it tells us we shouldn't just throw in every possible augmentation technique; there’s a sweet spot for data preparation when dealing with minute objects like bees.

Lalam: This summary shows that the paper isn't just about building a model; it’s about systematically dissecting the learning process to see exactly where those performance gains come from, which is valuable for building more reliable systems.

The paper's improvements: Tom: When we look at the suggested improvements from "Bee Detection and Tracking at Hive Entrance using YOLO11 and ByteTrack," they are pointing toward specific strategies that enhance the detection model, tracking algorithm, and the overall counting logic.

Jane: They suggest using a progressive backbone unfreezing strategy for training because it seems to give the most stable results, achieving about ninety-seven point zero percent precision and ninety-eight point seven percent mAP50 across their experiments while showing better convergence than just freezing everything or fine-tuning everything at once <ref:2608.23213#pg0,about 97.0% precision and 98.7% mAP50>.

Lu: That progressive freezing scheme is a clever way to mitigate overfitting while still allowing the network to learn the complex features needed for those tiny bees, which is a smart training methodology for this scenario.

Meng: From a practical standpoint, implementing that progressive unfreezing means we can get good results even when our dataset isn't massive, which is exactly what we need for practical applications like monitoring hives in the field.

Lalam: They also recommend tuning ByteTrack parameters quite specifically—like lowering the high and low thresholds and increasing the buffer size—to keep tracks alive longer under motion blur, showing that tracking parameter optimization is just as critical as the detector itself.

Conclusion: Tom: To wrap things up on "Bee Detection and Tracking at Hive Entrance using YOLO11 and ByteTrack," the paper concludes that a combination of moderate augmentation, progressive backbone unfreezing, and careful ByteTrack tuning significantly improves reliability under realistic recording conditions for this task.

Jane: So, the main implication is that by focusing on these specific optimization levers rather than just using the newest model out of the box, we can make automated monitoring much more trustworthy when dealing with fast-moving subjects.

Lu: The paper points out that their main limitation was still missing detections of fast-moving outgoing bees, suggesting that future work should focus on improving robustness specifically against viewpoint changes and lower frame rates.

Meng: That limitation is important because it tells us exactly where the current system stops working; it needs to be hardened against the exact scenarios where bees move too quickly between frames.

Lalam: I think this whole study, "Bee Detection and Tracking at Hive Entrance using YOLO11 and ByteTrack," shows that deep learning applied to ecology isn't just theoretical; when you balance the detector, the tracker, and your training data strategy correctly, you get a functional monitoring tool for colony dynamics.

Thi Thu Thao Nguyen, Johannes Reschke

Savonia University of Applied Sciences · Ostbayerische Technische Hochschule Regensburg

cs.CV

Submitted: 2026-08-24

Updated: 2026-10-04

Comments: 17 pages, 13 figures

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 75/100

The gist: This work presents an automatic bee entrance monitoring system based on YOLO11 transfer learning and ByteTrack tracking, investigating how data augmentation, backbone freezing, and tracker parameter

Key concepts

YOLO11: A state-of-the-art object detector used to find bees in videos. It was chosen over YOLOv8 because it offers better feature representation and localization, which is crucial for accurately spotting small, fast-moving insects like bees.
ByteTrack: A tracking algorithm that follows detected objects across consecutive video frames. It was tuned with specific parameters to handle bees that move quickly or have blurry detections by improving how it associates detections and keeps tracks alive during temporary occlusions.
Progressive Backbone Freezing: A training strategy where the model's deep layers are frozen initially, then gradually unfrozen during training. This method was found to provide the most stable performance and balanced learning, helping the model learn task-specific features without overfitting.
Data Augmentation Effects: The study tested different ways of artificially increasing the training data. Heavy augmentation hurt performance for small objects like bees, while light augmentation was better. This suggests a moderate approach to adding variations to the training images is optimal for bee detection.

Terminology

Accuracy — 24× in this paper · explained in 1 episode
Data Augmentation — 8× in this paper · explained in 7 episodes
Transfer Learning — 8× in this paper · explained in 10 episodes
Fine-Tuning — 5× in this paper · explained in 5 episodes
Robustness — 5× in this paper · explained in 5 episodes
Consistency — 3× in this paper · explained in 3 episodes
Generalization — 3× in this paper · explained in 13 episodes
Adaptability — 1× in this paper · explained in 1 episode
Core — 1× in this paper · explained in 2 episodes

Summary

This work presents an automatic bee entrance monitoring system based on YOLO11 transfer learning and ByteTrack tracking, investigating how data augmentation, backbone freezing, and tracker parameter optimization influence the detection and counting of small, fast-moving bees.

Detection Model Strategy

The study evaluates YOLOv8 and YOLO11 for bee detection. Both are onestage, anchor-free object detectors that provide high detection accuracy with real-time inference speed. YOLO11 is noted to offer improved feature representation and localization performance, making it more suitable for detecting small, fast-moving bees compared to YOLOv8. Both models were initialized with COCO pre-trained weights and fine-tuned on the bee dataset using transfer learning, which accelerates convergence, improves detection accuracy with limited data, and reduces the risk of overfitting.

Data Augmentation Effects

Experiments investigated how different augmentation strategies impacted model performance. The findings indicated that heavy augmentation with RandAugment and random erasing can brutally diminish model performance in terms of mAP50-95, box loss and class loss, and recall value for small objects like bees. Conversely, light augmentation consistently achieves higher localization accuracy, higher recall and lower losses, suggesting an optimization in the model’s learning process.

Backbone Freezing Schemes

The research compared three training strategies concerning backbone freezing:

Fully unfreezing all layers (fine-tune all).
Freezing the backbone for the first 30 epochs and then unfreezing (progressive backbone unfreezing).
Freezing the backbone throughout the entire training process.

The results showed that progressive backbone freezing delivers the most stable and balanced performance, achieving high mAP50 (ff 0.986–0.987), highest precision (ff0.971–0.974) and around 96% recall rate. This strategy was identified as providing the best stability and task-specific learning while mitigating overfitting, as it gained the lowest validation DFL loss (ff0.955).

Tracking Algorithm Optimization

The ByteTrack algorithm was employed to track detected bees across consecutive frames. It was selected because honey bees are small, fast-moving objects that often produce low-confidence detections due to motion blur and viewpoint changes. To improve trajectory continuity under challenging conditions, the tracker parameters were optimized:

((

(track high thresh = 0.15): Lower than default (0.25) to allow more valid bee detections into the first association stage.

(track low thresh = 0.03): Lower than default (0.10) to enable recovery of weak detections during the second association stage.

(new track thresh = 0.15): Lower than default (0.25) to initialize newly appearing bees more quickly.

(track buffer = 50): Higher than default (30) to keep lost tracks alive for a longer period, enabling successful re-association after temporary occlusion.

(match thresh = 0.85): Higher than default (0.80) to require stronger spatial consistency before assigning an existing ID, reducing incorrect ID switches.

Counting Logic and Error Analysis

The counting logic utilized a rectangular counting box defined at the hive entrance. A bee is counted as IN or OUT when its center point crosses the boundary of this box, with each tracked ID added to the IN/OUT log independently. The error analysis revealed that most counting errors were caused by missed detections due to rapid bee motion and motion blur, especially for outgoing bees, which move much faster than incoming bees. Tracking failures primarily occurred when bees moved quickly between consecutive frames, leading to large displacements that made association difficult.

Overall Conclusion

The study concludes that moderate augmentation, progressive backbone unfreezing, and ByteTrack tuning improve the reliability of automatic bee entrance monitoring under realistic recording conditions. The main limitation identified is the missed detection of fast-moving outgoing bees, suggesting future work should focus on improving robustness to viewpoint changes and low-frame-rate videos.

The gist

The optimized YOLO11-ByteTrack system achieved a 91.5% counting accuracy for incoming bees and a 23.3% counting accuracy for outgoing bees in a side-view video, demonstrating that moderate augmentation, progressive backbone unfreezing, and ByteTrack tuning improve the reliability of automatic bee entrance monitoring under realistic recording conditions.

How it works

Detection: YOLO11 is used as the detector, fine-tuned via transfer learning from COCO weights on the bee dataset to achieve high precision and localization accuracy.

Improvements for AI systems

Here are the specific, actionable improvements for an AI system based on this research, categorized by component:

) Detection Model Improvements:

Implement a tiered training strategy utilizing a Progressive Backbone Unfreezing schedule (freezing the backbone for initial epochs, then unfreezing gradually). This is crucial for small object detection stability under limited data conditions.
Adopt YOLOv11 architecture over YOLOv8 for detection tasks, leveraging its improved feature representation and localization performance specifically beneficial for detecting small, fast-moving objects like bees.
Integrate Light Data Augmentation strategies (e.g., controlled HSV adjustments and minor translations/scales) instead of Heavy Augmentation (like RandAugment/Random Erasing). This prevents the distortion of tiny bee features which often leads to catastrophic recall loss in small object detection.

) Tracking Algorithm Improvements:

Implement a highly tuned ByteTrack configuration with specific parameters derived from the study:

Bottleneck Parameter Tuning: Set low confidence thresholds (e.g., 0.03 for low-confidence detections) and a larger track buffer (e.g., 50) to maximize the recovery of temporary track losses caused by partial occlusion and rapid motion blur inherent in hive entrance environments.

Enhance feature extraction within the detection pipeline by incorporating attention mechanisms (like C2PSA or C3k2 blocks, as suggested in Experiment 3) into the YOLO11 backbone to improve feature consistency across consecutive frames, thereby reducing ID switches during tracking.

) Counting and System Logic Improvements:

Utilize a Counting Box approach over a single virtual counting line for ingress/egress quantification. This provides superior robustness against oscillatory flight behaviors near the entrance by calculating boundary crossings based on the center point trajectory relative to a fixed box perimeter, rather than relying solely on linear movement.

) System Capabilities and Outcomes:

The improved AI system can reliably perform the following:

Accurately count bee traffic (both entering and exiting) at hive entrances with high precision (up to 91.5% for incoming bees and significantly improved outgoing bee counting).
Maintain trajectory continuity for fast-moving subjects even under challenging conditions like motion blur and low frame rates, leading to a significant reduction in tracking failures compared to standard systems.
Provide robust monitoring of colony dynamics by accurately quantifying the rate and direction of bee movement, which is vital for assessing colony health or swarming behavior indicators.

Sources

YOLOv11: An Overview of the Key Architectural Enhancements