https://shadowscythe03.github.io/tags/cnns/CNNsHugo Blox Builder (https://hugoblox.com)en-usWed, 01 Nov 2023 00:00:00 +0000https://shadowscythe03.github.io/media/logo_hu_f3e4bd98189c9234.pnghttps://shadowscythe03.github.io/tags/cnns/https://shadowscythe03.github.io/projects/aerial-segmentation/Wed, 01 Nov 2023 00:00:00 +0000https://shadowscythe03.github.io/projects/aerial-segmentation/<p>A systematic benchmark of four segmentation architectures on the — 400 high-resolution aerial images (up to 4000×6000 px) with pixel-level labels across 23 urban classes. The goal: understand the accuracy–cost tradeoff across CNN and transformer approaches under real GPU memory constraints.</p> <p><strong>Best result: UNet at 86.4% accuracy</strong>, outperforming DeepLabV3+ (~80%), Swin Transformer (~82%), and SegNet (~75%).</p> <h2 id="dataset">Dataset</h2> <p><strong>Semantic Drone Dataset</strong> — 400 images, 350 train / 50 test, 23 semantic classes:</p> <ul> <li><strong>Structures</strong>: building, wall, fence, bridge</li> <li><strong>Nature</strong>: tree, vegetation, grass, water</li> <li><strong>Infrastructure</strong>: road, sidewalk, parking, rail track</li> <li><strong>Objects</strong>: car, truck, person, bicycle</li> <li><strong>Environment</strong>: sky, ground, terrain</li> </ul> <p>Images were downscaled to 384×512 for training due to GPU memory constraints (original 4000×6000 px).</p> <h2 id="architectures">Architectures</h2> <table> <thead> <tr> <th>Model</th> <th>Params</th> <th>Accuracy</th> <th>Inference</th> </tr> </thead> <tbody> <tr> <td><strong>UNet</strong></td> <td>~34.6M</td> <td><strong>86.4%</strong></td> <td>Fast</td> </tr> <tr> <td>Swin Transformer</td> <td>—</td> <td>~82%</td> <td>Slow</td> </tr> <tr> <td>DeepLabV3+ (ResNet-50)</td> <td>—</td> <td>~80%</td> <td>Medium</td> </tr> <tr> <td>SegNet (VGG)</td> <td>—</td> <td>~75%</td> <td>Fast</td> </tr> </tbody> </table> <p>All models trained with Adam (lr=0.001), sparse categorical crossentropy, up to 500 epochs with early stopping; convergence around epoch 300–400.</p> <p><strong>Key tradeoffs:</strong></p> <ul> <li>UNet&rsquo;s encoder–decoder with skip connections gave the best overall accuracy with fast inference.</li> <li>Swin Transformer captured finer details thanks to global attention but at high compute cost.</li> <li>SegNet was the most memory-efficient via pooling-index upsampling.</li> <li>DeepLabV3+ with ASPP is best suited for full-resolution inference when memory allows.</li> </ul> <p><strong>Known limitation:</strong> all models struggled with small objects (vehicles, pedestrians &lt;60% accuracy) due to downscaling; patch-based inference at full resolution would close this gap.</p> <h2 id="technical-stack">Technical Stack</h2> <p>TensorFlow 2.x / Keras · NumPy · OpenCV · Google Colab (T4 GPU) · TensorBoard</p> <h2 id="project-report">Project Report</h2> <!-- TODO: Upload your project report PDF and replace the src URL below. Recommended: upload as a GitHub Release asset on this repo, then use: https://docs.google.com/viewer?url=<raw-pdf-url>&embedded=true --> <iframe src="https://docs.google.com/viewer?url=https://raw.githubusercontent.com/shadowscythe03/projects/main/aerial_sem_seg_FINAL_SUBMISSION/6-final_report.pdf&embedded=true" width="100%" height="850px" style="border: 1px solid #e5e7eb; border-radius: 8px; display: block; margin: 1rem 0;"> </iframe> <p> </p> <hr> <p><strong>Course</strong>: CS5480 — Deep Learning for Vision, IIT Hyderabad (Aug–Nov 2023) · Supervised by Prof. Summohana Channapayya</p>