<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[Biswas Labs]]></title><description><![CDATA[Discover how I built an Apple Detection system using YOLOv8 and Computer Vision. Explore dataset preparation, model training, object detection, and deep learning in action.]]></description><link>https://bisw.hashnode.dev</link><generator>RSS for Node</generator><lastBuildDate>Sat, 21 Nov 2026 19:06:46 GMT</lastBuildDate><atom:link href="https://bisw.hashnode.dev/rss.xml" rel="self" type="application/rss+xml"/><language><![CDATA[en]]></language><ttl>60</ttl><item><title><![CDATA[Building an Apple Detection System Using YOLOv8]]></title><description><![CDATA[Build a real-time apple detection and counting system using YOLOv8, dataset engineering, INT8 quantization, and edge AI optimization.]]></description><link>https://bisw.hashnode.dev/building-an-apple-detection-system-using-yolov8</link><guid isPermaLink="true">https://bisw.hashnode.dev/building-an-apple-detection-system-using-yolov8</guid><category><![CDATA[Computer Vision]]></category><category><![CDATA[DeepLearning]]></category><category><![CDATA[object detection ]]></category><category><![CDATA[YOLO]]></category><category><![CDATA[ Edge AI]]></category><dc:creator><![CDATA[biswas pokhrel]]></dc:creator><pubDate>Mon, 09 Nov 2026 05:01:42 GMT</pubDate><enclosure url="https://cdn.hashnode.com/uploads/covers/6af028e0b1d5b30159db109b/594c8065-2065-43a5-b087-00b22cfa57b3.png" length="0" type="image/jpeg"/><content:encoded><![CDATA[<h2>Introduction</h2>
<p>Computer Vision has made it possible to build intelligent systems that can understand and interact with the physical world. From autonomous vehicles to industrial automation, object detection is becoming an essential component of modern AI applications.</p>
<p>One interesting application is agricultural automation.</p>
<p>Imagine a system capable of identifying apples in an orchard, counting fruits on a conveyor belt, or detecting individual apples for automated harvesting. These applications require more than simply recognizing an object in an image. They require reliable detection under different environmental conditions, including occlusion, overlapping objects, varying lighting, and different object densities.</p>
<p>This motivated me to build <strong>AppleVision-Edge</strong>, an apple detection and counting system using YOLOv8, with a focus on optimizing deep learning models for resource-constrained edge devices such as Raspberry Pi.</p>
<p>The project explores the complete computer vision pipeline, starting from dataset preparation and model training to model optimization, quantization, and deployment-oriented inference.</p>
<p><strong>Project Repository:</strong> <a href="https://github.com/biswas-bit/apple-detection-edge">AppleVision-Edge</a></p>
<p><img src="https://github.com/biswas-bit/apple-detection-edge/raw/main/apple-detection-edge/testimg/Screenshot%202026-09-05%20134153.png" alt="AppleVision-Edge Detection Result" /></p>
<hr />
<h2>1. The Problem: Why Apple Detection?</h2>
<p>Detecting apples might initially appear to be a straightforward object detection problem. However, real-world agricultural environments introduce several challenges.</p>
<p>Consider an orchard containing hundreds of apples. Some fruits may be partially hidden behind leaves, while others overlap with neighboring apples. Changes in lighting, camera distance, and apple density further complicate detection.</p>
<p>During development, I identified several challenges:</p>
<ul>
<li><strong>Occlusion:</strong> Apples are frequently partially hidden behind leaves or other fruits.</li>
<li><strong>Object clustering:</strong> Closely packed apples make individual object localization difficult.</li>
<li><strong>Domain variation:</strong> Models trained on isolated fruit images may struggle with orchard environments.</li>
<li><strong>Deployment constraints:</strong> Deep learning models must be optimized to operate on hardware with limited computational resources.</li>
</ul>
<p>The objective was therefore not simply to achieve a high detection score on a particular dataset. It was to investigate how a detection model behaves across different imaging conditions and prepare it for edge deployment.</p>
<h2>2. Why YOLOv8?</h2>
<p>For this project, I selected YOLOv8, developed by Ultralytics.</p>
<p>YOLO (You Only Look Once) belongs to the family of single-stage object detectors. Unlike traditional two-stage detection approaches, which first generate region proposals and then classify them, YOLO performs object detection through a unified prediction pipeline.</p>
<p>This makes YOLO particularly attractive for applications where inference speed matters.</p>
<p>I selected the lightweight YOLOv8n (Nano) variant because the project targets resource-constrained hardware.</p>
<p>The initial model was initialized using COCO-pretrained weights and fine-tuned for a single object class: apple.</p>
<p>The main reasons for choosing YOLOv8 were:</p>
<ul>
<li>Lightweight model variants.</li>
<li>Transfer learning using pretrained weights.</li>
<li>Support for custom object detection datasets.</li>
<li>Convenient model export options.</li>
<li>An established ecosystem for training, evaluation, and inference.</li>
</ul>
<h2>3. Dataset Engineering: The Most Important Part</h2>
<p>One of the most important lessons from this project was that model performance depends heavily on the quality and diversity of the training data.</p>
<p>Instead of relying on a single dataset, I experimented with multiple apple datasets representing different imaging conditions.</p>
<h3>Datasets Used</h3>
<p><strong>1. Isolated Apple Dataset</strong></p>
<p>This dataset primarily contains clean images with isolated or relatively few apples.</p>
<p>It was useful for establishing an initial baseline and understanding how well the model could learn the target object.</p>
<p><strong>2. MinneApple Dataset</strong></p>
<p>MinneApple provides orchard images containing apples under more realistic agricultural conditions.</p>
<p>Unlike isolated fruit images, these scenes introduce challenges such as overlapping fruits, foliage, and heavy occlusion.</p>
<p>The dataset originally provides instance segmentation annotations, which I converted into YOLO-compatible bounding-box annotations.</p>
<p>Research reference: Häni and Roy, <em>MinneApple: A Benchmark Dataset for Apple Detection and Segmentation</em> (2019).</p>
<p><strong>3. Fuji Apple Dataset</strong></p>
<p>The Fuji dataset provides close-up apple imagery and introduces another imaging domain.</p>
<p>Its original VIA polygon CSV annotations were also converted into a unified YOLO bounding-box format.</p>
<h3>Annotation Unification</h3>
<p>A major challenge was that the datasets did not share the same annotation format.</p>
<table>
<thead>
<tr>
<th>Dataset</th>
<th>Original Annotation Format</th>
</tr>
</thead>
<tbody><tr>
<td>Isolated Apple</td>
<td>YOLO bounding boxes</td>
</tr>
<tr>
<td>MinneApple</td>
<td>Instance segmentation masks</td>
</tr>
<tr>
<td>Fuji</td>
<td>VIA polygon CSV</td>
</tr>
</tbody></table>
<p>I developed a companion dataset engineering pipeline to convert these different annotation formats into a unified YOLO detection format.</p>
<p>This allowed the datasets to be used within a consistent training workflow.</p>
<p><strong>Dataset Pipeline:</strong> <a href="https://github.com/biswas-bit/apple-data-unification">apple-dataset-pipeline</a></p>
<h2>4. Training the YOLOv8 Model</h2>
<p>After preparing the datasets, I fine-tuned the COCO-pretrained YOLOv8n model for apple detection.</p>
<p>The training pipeline used a 320 × 320 input resolution.</p>
<p>This resolution was selected as an initial trade-off between computational requirements and detection accuracy, considering the eventual Raspberry Pi deployment target.</p>
<p>The main training configuration included:</p>
<ul>
<li>Architecture: YOLOv8n</li>
<li>Initialization: COCO-pretrained weights</li>
<li>Detection classes: 1 (Apple)</li>
<li>Input resolution: 320 × 320</li>
<li>Training duration: Up to 150 epochs</li>
<li>Bounding-box optimization experiments: CIoU and GIoU</li>
</ul>
<h3>Understanding the Importance of Generalization</h3>
<p>My initial experiment produced an impressive <a href="mailto:mAP@0.5">mAP@0.5</a> score of 0.993.</p>
<p>However, testing the model on crowded and occluded apple images revealed an important weakness: the model did not generalize well to more complex scenes.</p>
<p>This was a valuable lesson.</p>
<p>A high evaluation score does not necessarily mean that a model has learned the underlying problem effectively. Dataset homogeneity and limited scene diversity can produce misleadingly optimistic results.</p>
<p>To investigate this issue, I moved toward more diverse orchard imagery, particularly the MinneApple dataset.</p>
<p>This shifted the focus from achieving a high score on a relatively simple dataset to improving performance under realistic detection conditions.</p>
<h2>5. Performance Benchmarking: Accuracy vs. Efficiency</h2>
<p>A detection model is not evaluated solely by how accurately it identifies objects.</p>
<p>For practical deployment, its computational complexity, memory consumption, model size, and inference speed are equally important.</p>
<p>Since AppleVision-Edge targets resource-constrained hardware, I evaluated the trained models using multiple performance metrics.</p>
<p>The experiments compare three model configurations:</p>
<ul>
<li><strong>MinneApple:</strong> Model trained using the MinneApple dataset.</li>
<li><strong>Fuji:</strong> Model trained using the Fuji apple dataset.</li>
<li><strong>Merged:</strong> Model trained using the merged dataset configuration and exported in INT8 TensorFlow Lite format.</li>
</ul>
<p>The following sections examine their performance from different perspectives.</p>
<h3>5.1 Model Size and Computational Complexity</h3>
<p>To evaluate deployment efficiency, I measured the trainable parameter count, model size, and computational complexity in GFLOPs.</p>
<p>These metrics provide insight into the model's storage requirements and computational demands.</p>
<table>
<thead>
<tr>
<th>Model</th>
<th align="right">Trainable Parameters</th>
<th align="right">Model Size (MB)</th>
<th align="right">GFLOPs</th>
</tr>
</thead>
<tbody><tr>
<td>MinneApple</td>
<td align="right">3,011,043</td>
<td align="right">5.92</td>
<td align="right">8.200</td>
</tr>
<tr>
<td>Fuji</td>
<td align="right">3,011,043</td>
<td align="right">5.93</td>
<td align="right">8.200</td>
</tr>
<tr>
<td>Merged</td>
<td align="right">3,012,240</td>
<td align="right">3.15</td>
<td align="right">2.048</td>
</tr>
</tbody></table>
<p>One interesting observation is that the merged model contains a similar number of trainable parameters to the other two models, yet its reported model size and computational complexity are substantially lower.</p>
<p>The merged model has approximately 2.05 GFLOPs compared with 8.2 GFLOPs for the individual models.</p>
<p>This represents a significant reduction in reported computational complexity.</p>
<p>It is important to distinguish parameter count from computational complexity. Having a similar number of parameters does not necessarily mean that two models require the same amount of computation during inference.</p>
<h3>5.2 Detection Accuracy</h3>
<p>Model efficiency is useful only when the model can maintain acceptable detection performance.</p>
<p>To investigate this trade-off, I compared the models using two important object detection metrics:</p>
<ul>
<li><strong><a href="mailto:mAP@0.5">mAP@0.5</a>:</strong> Mean Average Precision at an IoU threshold of 0.5.</li>
<li><strong><a href="mailto:mAP@0.5">mAP@0.5</a>:0.95:</strong> Mean Average Precision averaged across multiple IoU thresholds from 0.5 to 0.95.</li>
</ul>
<p>I also recorded Precision, Recall, and F1-score for the merged model.</p>
<table>
<thead>
<tr>
<th>Model</th>
<th align="right"><a href="mailto:mAP@0.5">mAP@0.5</a></th>
<th align="right"><a href="mailto:mAP@0.5">mAP@0.5</a>:0.95</th>
<th align="right">Precision</th>
<th align="right">Recall</th>
<th align="right">F1-score</th>
</tr>
</thead>
<tbody><tr>
<td>MinneApple</td>
<td align="right">0.65</td>
<td align="right">0.26</td>
<td align="right">—</td>
<td align="right">—</td>
<td align="right">—</td>
</tr>
<tr>
<td>Fuji</td>
<td align="right"><strong>0.83</strong></td>
<td align="right"><strong>0.60</strong></td>
<td align="right">—</td>
<td align="right">—</td>
<td align="right">—</td>
</tr>
<tr>
<td>Merged</td>
<td align="right">0.69</td>
<td align="right">0.31</td>
<td align="right"><strong>0.78</strong></td>
<td align="right"><strong>0.62</strong></td>
<td align="right"><strong>0.6953</strong></td>
</tr>
</tbody></table>
<h3>Observations</h3>
<p><strong>Fuji Model</strong></p>
<p>The Fuji-trained model achieved the highest detection accuracy, with a <a href="mailto:mAP@0.5">mAP@0.5</a> of 0.83 and <a href="mailto:mAP@0.5">mAP@0.5</a>:0.95 of 0.60.</p>
<p>This indicates stronger detection and localization performance on its evaluated dataset.</p>
<p><strong>MinneApple Model</strong></p>
<p>The MinneApple model achieved a <a href="mailto:mAP@0.5">mAP@0.5</a> of 0.65.</p>
<p>The orchard environment introduces additional challenges, particularly object clustering and occlusion.</p>
<p><strong>Merged Model</strong></p>
<p>The merged model achieved a <a href="mailto:mAP@0.5">mAP@0.5</a> of 0.69 and <a href="mailto:mAP@0.5">mAP@0.5</a>:0.95 of 0.31.</p>
<p>Its precision of 0.78 and recall of 0.62 indicate that it identifies a reasonable proportion of the evaluated apple instances, although some objects remain undetected.</p>
<p>The F1-score of 0.6953 provides a combined measure of precision and recall.</p>
<p>The important observation is that the merged model does not achieve the highest detection accuracy. Instead, it provides a different balance between detection performance and computational efficiency.</p>
<h3>5.3 Inference Latency</h3>
<p>For an edge deployment scenario, inference speed is another critical factor.</p>
<p>A model may achieve excellent detection accuracy but still be unsuitable for real-time applications if its inference latency is too high.</p>
<p>I measured the inference pipeline using three latency components:</p>
<ul>
<li><strong>Pre-processing:</strong> Preparing and resizing the input image.</li>
<li><strong>Inference:</strong> Running the neural network to generate predictions.</li>
<li><strong>Post-processing:</strong> Processing predictions and producing final detections.</li>
</ul>
<p>The end-to-end latency represents the reported total processing time.</p>
<table>
<thead>
<tr>
<th>Model</th>
<th align="right">Pre-processing (ms)</th>
<th align="right">Inference (ms)</th>
<th align="right">Post-processing (ms)</th>
<th align="right">End-to-End (ms)</th>
<th align="right">FPS</th>
</tr>
</thead>
<tbody><tr>
<td>MinneApple</td>
<td align="right">0.88</td>
<td align="right">52.47</td>
<td align="right">1.29</td>
<td align="right">55.02</td>
<td align="right">18.18</td>
</tr>
<tr>
<td>Fuji</td>
<td align="right">1.06</td>
<td align="right">64.87</td>
<td align="right">1.58</td>
<td align="right">67.96</td>
<td align="right">14.12</td>
</tr>
<tr>
<td>Merged</td>
<td align="right">1.10</td>
<td align="right">34.51</td>
<td align="right">1.63</td>
<td align="right"><strong>37.81</strong></td>
<td align="right"><strong>26.45</strong></td>
</tr>
</tbody></table>
<h3>What Do These Results Tell Us?</h3>
<p>The merged model achieved the lowest reported inference latency at 34.51 ms.</p>
<p>Its reported end-to-end latency was 37.81 ms, corresponding to approximately 26.45 FPS.</p>
<p>Compared with the other configurations, this represents a considerable improvement in inference speed.</p>
<p>The Fuji model achieved the highest detection accuracy but had the highest inference latency.</p>
<p>The merged model, on the other hand, achieved faster inference while maintaining a <a href="mailto:mAP@0.5">mAP@0.5</a> of 0.69.</p>
<p>This demonstrates an important trade-off in computer vision deployment: the most accurate model is not always the most suitable model for a resource-constrained environment.</p>
<h3>5.4 RAM Consumption</h3>
<p>Model size alone does not tell us how much memory a model requires during inference.</p>
<p>A model also needs memory for loading weights, intermediate computations, input tensors, and other runtime operations.</p>
<p>To understand this, I measured model loading memory, inference overhead, and total memory consumption.</p>
<table>
<thead>
<tr>
<th>Model</th>
<th>Format</th>
<th align="right">Model Load (MB)</th>
<th align="right">Inference Overhead (MB)</th>
<th align="right">Total Memory (MB)</th>
</tr>
</thead>
<tbody><tr>
<td>MinneApple</td>
<td>FP32 (.pt)</td>
<td align="right">19.78</td>
<td align="right">28.07</td>
<td align="right">47.86</td>
</tr>
<tr>
<td>Fuji</td>
<td>FP32 (.pt)</td>
<td align="right">8.58</td>
<td align="right">8.01</td>
<td align="right">16.60</td>
</tr>
<tr>
<td>Merged</td>
<td>INT8 (TFLite)</td>
<td align="right">0.02</td>
<td align="right">4.55</td>
<td align="right"><strong>4.58</strong></td>
</tr>
</tbody></table>
<p>The merged INT8 model demonstrated a substantial reduction in reported memory consumption.</p>
<p>Its total measured memory usage was approximately 4.58 MB, compared with 47.86 MB for the MinneApple FP32 model.</p>
<p>This is particularly interesting for edge computing, where memory availability can be a major limitation.</p>
<p>The use of INT8 quantization reduces numerical precision from 32-bit floating-point representation to 8-bit integer representation, which can reduce storage and memory requirements.</p>
<p>However, actual memory usage also depends on the inference runtime, tensor allocation, and hardware environment.</p>
<h3>5.5 Overall Benchmark Analysis</h3>
<p>The experiments highlight an important relationship between detection accuracy and deployment efficiency.</p>
<table>
<thead>
<tr>
<th>Evaluation Criteria</th>
<th>Best Performing Model</th>
</tr>
</thead>
<tbody><tr>
<td><a href="mailto:mAP@0.5">mAP@0.5</a></td>
<td>Fuji</td>
</tr>
<tr>
<td><a href="mailto:mAP@0.5">mAP@0.5</a>:0.95</td>
<td>Fuji</td>
</tr>
<tr>
<td>Model Size</td>
<td>Merged</td>
</tr>
<tr>
<td>GFLOPs</td>
<td>Merged</td>
</tr>
<tr>
<td>Inference Latency</td>
<td>Merged</td>
</tr>
<tr>
<td>Reported FPS</td>
<td>Merged</td>
</tr>
<tr>
<td>RAM Consumption</td>
<td>Merged</td>
</tr>
</tbody></table>
<p>The Fuji model achieved the strongest detection accuracy, while the merged INT8 model demonstrated better computational and memory efficiency.</p>
<p>For applications requiring maximum detection accuracy, the Fuji configuration may be preferable on its evaluated dataset.</p>
<p>However, for resource-constrained deployment scenarios, the merged model presents a more attractive balance between accuracy, speed, and memory consumption.</p>
<p>This is the central engineering trade-off explored in AppleVision-Edge.</p>
<h2>6. From Object Detection to Apple Counting</h2>
<p>Object detection provides bounding boxes and confidence scores, but many agricultural applications require more than identifying objects.</p>
<p>They require counting them.</p>
<p>I explored multiple counting strategies within AppleVision-Edge.</p>
<h3>Per-frame Detection</h3>
<p>The simplest approach counts the number of detected apples in an individual image or video frame.</p>
<p>This is useful for static images and frame-level analysis.</p>
<h3>ByteTrack-based Tracking</h3>
<p>For video applications, counting detections independently in every frame can result in repeatedly counting the same apple.</p>
<p>To address this, I explored ByteTrack-based object tracking.</p>
<p>Tracking associates detections across consecutive frames, allowing the system to maintain object identities over time.</p>
<h3>Line-crossing Counting</h3>
<p>For conveyor-belt applications, a virtual counting line can be introduced.</p>
<p>When a tracked apple crosses the line, the system increments the cumulative count.</p>
<p>This approach is particularly relevant to automated fruit sorting and packaging systems.</p>
<h2>7. Preparing the Model for Edge Deployment</h2>
<p>Training a model in a GPU-powered environment is different from running it on a small edge device.</p>
<p>A Raspberry Pi has considerably more limited computational resources than a typical machine learning development workstation.</p>
<p>Consequently, deploying the original model without optimization may not provide the desired inference performance.</p>
<p>For this reason, I explored TensorFlow Lite export with INT8 post-training quantization.</p>
<h3>What Is INT8 Quantization?</h3>
<p>Neural networks commonly perform computations using 32-bit floating-point numbers (FP32).</p>
<p>Quantization converts model weights and, depending on the quantization approach, activations into lower-precision representations.</p>
<p>INT8 quantization represents values using 8-bit integers.</p>
<p>This can provide several benefits:</p>
<ul>
<li>Reduced model storage requirements.</li>
<li>Lower memory consumption.</li>
<li>Potentially faster inference on compatible hardware.</li>
<li>Better suitability for resource-constrained devices.</li>
</ul>
<p>For AppleVision-Edge, the merged model was exported to TensorFlow Lite with INT8 quantization.</p>
<p>The intended inference runtime is <code>tflite-runtime</code>, with the XNNPACK delegate for CPU optimization.</p>
<h3>Target Deployment Configuration</h3>
<table>
<thead>
<tr>
<th>Component</th>
<th>Configuration</th>
</tr>
</thead>
<tbody><tr>
<td>Target Hardware</td>
<td>Raspberry Pi</td>
</tr>
<tr>
<td>Model</td>
<td>YOLOv8n</td>
</tr>
<tr>
<td>Input Resolution</td>
<td>320 × 320</td>
</tr>
<tr>
<td>Export Format</td>
<td>TensorFlow Lite</td>
</tr>
<tr>
<td>Quantization</td>
<td>INT8</td>
</tr>
<tr>
<td>Inference Runtime</td>
<td>tflite-runtime</td>
</tr>
<tr>
<td>CPU Optimization</td>
<td>XNNPACK</td>
</tr>
</tbody></table>
<p>The benchmark results demonstrate the efficiency of the merged INT8 model in the measured environment. Physical Raspberry Pi latency and FPS benchmarking remain part of the project's future validation work.</p>
<h2>8. Limitations and Future Improvements</h2>
<p>Although AppleVision-Edge explores the detection and deployment workflow, there are several areas that require further investigation.</p>
<h3>Current Limitations</h3>
<ul>
<li>Detection accuracy decreases in highly crowded and occluded scenes.</li>
<li>Generalization to completely unseen apple varieties and environmental conditions is not guaranteed.</li>
<li>Precise localization for robotic picking remains challenging.</li>
<li>Depth estimation and 3D spatial understanding are not currently implemented.</li>
<li>Physical Raspberry Pi latency and FPS measurements are still pending.</li>
</ul>
<h3>Future Improvements</h3>
<ol>
<li>Experiment with higher-resolution training to improve bounding-box localization.</li>
<li>Expand dataset coverage with drone imagery and packing-line environments.</li>
<li>Benchmark FP32 and INT8 models on physical Raspberry Pi hardware.</li>
<li>Investigate depth estimation for robotic picking applications.</li>
<li>Further optimize the inference pipeline for real-time performance.</li>
</ol>
<h2>Conclusion</h2>
<p>AppleVision-Edge started as an object detection project, but it gradually became an exploration of a much broader computer vision engineering workflow.</p>
<p>Working with multiple datasets highlighted the importance of data diversity and generalization. Comparing CIoU and GIoU provided practical insight into bounding-box optimization, while benchmarking different model configurations revealed the trade-offs between accuracy, computational complexity, inference speed, and memory consumption.</p>
<p>One of the most interesting findings was that the model with the highest detection accuracy was not necessarily the most efficient model for deployment.</p>
<p>The Fuji model achieved the highest detection accuracy, while the merged INT8 model demonstrated significantly lower computational requirements, faster inference, and reduced memory consumption.</p>
<p>Perhaps the most important lesson I learned was that achieving a high evaluation score is only one part of building a reliable machine learning system.</p>
<p>A useful model must also generalize to different environments, operate within hardware constraints, and solve an actual problem.</p>
<p>There is still work to do, particularly around physical edge-device benchmarking and improving detection in dense orchard scenes. However, this project has been a valuable step in understanding how computer vision models move from experimentation toward practical deployment.</p>
<p><strong>Explore the complete project:</strong></p>
<ul>
<li><a href="https://github.com/biswas-bit/apple-detection-edge">AppleVision-Edge GitHub Repository</a></li>
<li><a href="https://github.com/biswas-bit/apple-data-unification">Apple Dataset Pipeline</a></li>
<li><a href="https://arxiv.org/abs/1909.06441">MinneApple Research Paper</a></li>
<li><a href="https://docs.ultralytics.com">Ultralytics YOLOv8 Documentation</a></li>
</ul>
<hr />
<p><em>Built and documented by Biswas Pokhrel | Biswas Labs</em></p>
<p><strong>Exploring AI, One Project at a Time.</strong></p>
]]></content:encoded></item></channel></rss>