Hugging Face's logo

Buckets:
hamimmahmud0
/
thesis-presentation

Files
xet
hamimmahmud0/thesis-presentation / slides.md
hamimmahmud0's picture
hamimmahmud0
preview code
|
download
raw
48.8 kB
---
theme: default
title: Week 11 — Incremental Progress Update
info: |
Converted from “Week 6 Template (1).pptx”.
Original slide order, wording, and images are preserved.
author: Hamim Mahmud
canvasWidth: 1920
aspectRatio: 16/9
transition: slide-left
fonts:
sans: Ruda
weights: '400,500,600,700,800,900'
provider: google
mdc: true
---
<!-- Slide 1 -->
<div class="absolute inset-0 flex">
<div class="flex w-1/2 flex-col justify-center pl-20">
<div class="text-6xl font-black">CE 400</div>
<div class="mt-6 text-4xl font-bold">Hamim Mahmud</div>
<div class="text-3xl text-slate-400">2104065</div>
<div class="mt-10 max-w-xl font-black text-5xl leading-relaxed">
Incremental Progress
</div>
<div class="text-3xl text-slate-300">Aug 23, 2026</div>
</div>
<img class="w-1/2 object-cover grayscale" src="/assets/image2.png" alt="Traffic scene" />
</div>
---
<!-- Slide 2 -->
<div class="flex h-full flex-col justify-center pl-12">
<div class="text-6xl font-bold pb-10">Week progress</div>
<ul class="mt-4 list-disc space-y-4 pl-8 text-4xl leading-relaxed">
<li>Acquired Drone</li>
<li>Captured preliminary drone footage for trajectory<br>analysis pipeline</li>
<li>Took preliminary drone footage for photogrammetry</li>
<li>Application for acquiring CCTV footage permission from<br>DMP</li>
<li>Software selection for photogrammetry - (ongoing)</li>
</ul>
</div>
---
<!-- Slide 2.5 -->
<div class="flex h-full flex-col justify-center pl-12">
<div class="text-6xl font-bold pb-10">Week progress</div>
<ul class="mt-4 list-disc space-y-4 pl-8 text-4xl leading-relaxed">
<li>Vehicle Trajectory Extraction - (ongoing)
<ul>
<li>Video Stabilization</li>
<li>Vehicle segmentation</li>
<li>Orientation Extraction</li>
</ul>
</li>
<li>Infrastructure Development
<ul>
<li>Tunneling pipeline</li>
<li>SAM 3.1 MCP server</li>
<li>Cotracker MCP server</li>
<li>Integration with kaggle for Agentic Workflow</li>
</ul>
</li>
</ul>
</div>
---
<!-- Slide 3 -->
<div class="flex h-full items-center gap-16">
<ul class="list-disc space-y-4 pl-8 text-5xl">
<li>Model: DJI mini 2 (used)</li>
<li>Provider: DJI</li>
<li>Total cost: 32600 taka</li>
</ul>
<img class="h-4/5 object-contain rotate-90" src="/assets/image1.jpg" alt="DJI Mini 2, battery and controller" />
</div>
---
<!-- Slide 5 -->
<div class="flex h-full flex-col justify-center gap-10">
<h1 class="text-8xl font-black">Brief</h1>
<div class="text-4xl leading-relaxed">The science of extracting 3D geometric information from 2D<br>images</div>
<img class="mx-auto h-2/3 object-contain" src="/assets/image3.png" alt="Photogrammetry reconstruction workflow" />
</div>
---
<!-- Slide 6 -->
<div class="flex h-full flex-col justify-center gap-14">
<h1 class="text-5xl font-bold">Photogrammetry</h1>
<div class="flex justify-center gap-10">
<img class="h-128 object-cover" src="/assets/image4.png" alt="Polashi Intersection aerial frame 1" />
<img class="h-128 object-cover" src="/assets/image6.png" alt="Polashi Intersection aerial frame 2" />
</div>
<ul class="list-disc space-y-2 pl-8 text-2xl">
<li>Subject: Polashi Intersection</li>
<li>Captured amount: 151</li>
</ul>
<div class="text-lg text-slate-500">This images are only for selecting pipeline for photogrammetry</div>
</div>
---
<!-- Slide 7 -->
<div class="flex h-full flex-col justify-center gap-14">
<h1 class="text-9xl font-black">Existing Software</h1>
<div class="grid grid-cols-2 gap-16 text-2xl">
<div>
<strong class="text-4xl">Free:</strong>
<ul class="mt-4 list-disc space-y-2 pl-8 text-3xl">
<li>Meshroom</li>
<li>COLMAP</li>
<li>WebODM</li>
</ul>
</div>
<div>
<strong class="text-4xl">Paid:</strong>
<ul class="mt-4 list-disc space-y-2 pl-8 text-3xl">
<li>Agisoft Metashape</li>
<li>RealityCapture</li>
<li>3DF Zephyr</li>
<li>Bentley iTwin Capture<br>Modeler</li>
</ul>
</div>
</div>
<div class="text-4xl font-bold pl-4">Current preference: COLMAP</div>
</div>
---
<!-- Slide 8 -->
<div class="absolute inset-0 pt-10">
<h1 class="pl-10 font-black">Point Cloud</h1>
<img class="h-full w-full object-contain" src="/assets/image7.png" alt="Photogrammetry software screenshot" />
</div>
---
<!-- Slide 4 -->
<div class="flex h-full flex-col items-center justify-center gap-8">
<h1 class="text-7xl font-bold">Trajectory Analysis &amp; Surrogate Safety</h1>
<img class="h-3/1 object-contain" src="/assets/image5.gif" alt="Drone footage of Bakshi Bazaar" />
<div class="text-2xl">Location: Bakshi Bazaar</div>
</div>
---
<!-- Slide 10 -->
<div class="flex h-full flex-col justify-center gap-6 px-12">
<div>
<h1 class="text-5xl font-bold">Keypoint Tracking — CoTracker3</h1>
<div class="mt-2 text-2xl text-slate-300 font-medium">Estimating camera motion from the raw drone footage</div>
</div>
<img class="mx-auto h-2/1 object-contain" src="/assets/cotracker-keypoints.png" alt="Dense CoTracker3 keypoint grid over drone footage" />
<ul class="list-disc space-y-1 pl-8 text-3xl">
<li>Dense 16&times;16 grid — 32,400 tracked points per frame</li>
<li>7,856 frames (≈ 5.5 min) at 3840&times;2160, 23.976 fps</li>
<li>Point tracks are the input for both stabilization and motion analysis</li>
</ul>
</div>
---
<!-- Slide 11 -->
<div class="flex h-full flex-col justify-center gap-6 px-12">
<div>
<h1 class="text-5xl font-bold">Camera Motion Analysis</h1>
<div class="mt-2 text-xl text-slate-500">Derived from the CoTracker3 tracks</div>
</div>
<img class="mx-auto max-h-[88%] w-full object-contain" src="/assets/camera-motion-graphs.png" alt="Camera trajectory, translation, yaw, zoom scale and diagnostics plots" />
<div class="grid grid-cols-4 gap-6 text-2xl">
<div><strong>Path length</strong><br>3,280 px</div>
<div><strong>Mean speed</strong><br>0.418 px/frame</div>
<div><strong>Total yaw</strong><br>+0.74&deg;</div>
<div><strong>Net scale</strong><br>&times;1.0017</div>
</div>
<div class="text-lg text-slate-500">Drift is small but continuous — hover is not stable enough for direct trajectory extraction</div>
</div>
---
<!-- Slide 12 -->
<div class="flex h-full items-center gap-12 px-12">
<div class="flex w-1/2 flex-col gap-6">
<h1 class="text-5xl font-bold">Video Stabilization</h1>
<ul class="list-disc space-y-3 pl-8 text-2xl leading-relaxed">
<li>Per-frame homography fitted from the tracked points</li>
<li>Camera drift, yaw and zoom compensated back to the frame-0 reference</li>
<li>Result: a fixed world frame, so pixel coordinates map consistently across the clip</li>
<li>Prerequisite for metric trajectory extraction</li>
</ul>
<div class="text-lg text-slate-500">Preview: Polashi Intersection, stabilized output</div>
</div>
<video class="h-4/5 w-1/2 object-contain" src="/assets/stabilized-preview.mp4" autoplay loop muted playsinline></video>
</div>
---
<!-- Slide 13 -->
<div class="flex h-full flex-col justify-center gap-6 px-12">
<div>
<h1 class="text-5xl font-bold">Vehicle Segmentation — Meta SAM</h1>
<div class="mt-2 text-xl text-slate-500">Instance masks on the stabilized frames</div>
</div>
<img class="mx-auto h-3/5 object-contain" src="/assets/sam-segmentation.jpg" alt="SAM segmentation masks over an aerial intersection frame" />
<ul class="list-disc space-y-1 pl-8 text-2xl">
<li>Prompt-driven masks for vehicles, road surface and vegetation</li>
<li>Mask centroid and principal axis give position and orientation per vehicle</li>
<li>Next step: linking masks across frames into continuous trajectories</li>
</ul>
</div>
---
<!-- Slide 14 -->
<div class="flex h-full flex-col justify-center gap-6 px-12">
<div>
<h1 class="text-5xl font-bold">Infrastructure — Tunneling Pipeline</h1>
<div class="mt-2 text-xl text-slate-500">TunnelMate: exposing GPU-side model servers to the agent</div>
</div>
<img class="mx-auto h-3/2 object-contain" src="/assets/tunnelmate-dashboard.png" alt="TunnelMate operations dashboard" />
<ul class="list-disc space-y-1 pl-8 text-2xl">
<li>SAM 3.1 and CoTracker run as MCP servers on Kaggle GPU sessions</li>
<li>Tunnels publish those endpoints; the broker tracks state, traffic and expiry</li>
<li>Removes manual re-hosting every time a session restarts</li>
</ul>
</div>
---
<!-- Slide 15 -->
<div class="flex h-full items-center gap-12 px-12">
<div class="flex w-2/5 flex-col gap-6">
<h1 class="text-5xl font-black">Agentic Workflow</h1>
<ul class="list-disc space-y-3 pl-8 text-3xl leading-relaxed">
<li>OpenCode drives the MCP servers end to end — no manual step per clip</li>
<li>Health checks, job queueing and restart handled by the agent</li>
<li>Batch in progress: 125 images &times; 7 prompts</li>
<li>Frees the pipeline to run unattended between sessions</li>
</ul>
</div>
<video class="w-3/5 object-contain" src="/assets/opencode-workflow.mp4" autoplay loop muted playsinline></video>
</div>
---
<!-- Slide 16 -->
<div class="flex h-full items-center justify-center text-8xl font-black">Thank You</div>
---
class: week11-light
---
<!-- Week 11 — Cover -->
<div class="absolute inset-0 flex">
<div class="flex w-1/2 flex-col justify-center pl-20">
<div class="text-6xl font-black">CE 400</div>
<div class="mt-6 text-4xl font-bold">Hamim Mahmud</div>
<div class="text-3xl text-slate-400">2104065</div>
<div class="mt-10 max-w-xl font-black text-5xl leading-relaxed">
Incremental Progress
</div>
<div class="mt-4 text-3xl font-bold">Week 11</div>
<div class="text-3xl text-slate-300">Sep 6, 2026</div>
</div>
<img class="w-1/2 object-cover grayscale" src="/assets/image2.png" alt="Traffic scene" />
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Research focus</div>
<h1>Week 11 — Research Focus</h1>
<div class="cols"><div><p style="font-size:40px;font-weight:700">3D vehicle detection and world-location extraction from roadside cameras</p><p style="margin-top:32px">Benchmarking BEV models for Dhaka City</p><div class="note" style="margin-top:32px">Local road reconstruction supports the experiments. The research outcome is evidence about model performance in Dhaka.</div></div><div><img src="/assets/research-overview/drone-frame.jpg" alt="Own stabilized drone footage — local traffic scene"/><div class="caption">Own stabilized drone footage — local traffic scene</div></div></div>
</div>
---
class: week11-light
---
<!-- Week 11 — Progress -->
<div class="flex h-full flex-col justify-center pl-12">
<div class="text-6xl font-bold pb-10">Week 11 Progress</div>
<ul class="mt-4 list-disc space-y-4 pl-8 text-4xl leading-relaxed">
<li>Captured and stabilized two new drone videos</li>
<li>Fixed PLY point cloud color display in Blender<br>for the Gaussian splatting workflow</li>
<li>Conducted preliminary tests on previously acquired<br>photogrammetry data under different weather conditions</li>
<li>Identified the need to recapture data and improve<br>point cloud density</li>
</ul>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Research motivation</div>
<h1>How the earlier safety topic connects</h1>
<div class="flow"><div class="card"><h2>Earlier question</h2><p>Can CCTV footage reveal vehicle conflicts?</p></div><div class="arrow">→</div><div class="card"><h2>Required measurements</h2><p>Where are road users, and how do their paths change?</p></div><div class="arrow">→</div><div class="card"><h2>Current thesis</h2><p>Which BEV models estimate these spatial relationships reliably in Dhaka?</p></div></div><div class="cols"><div><img src="/assets/sam-segmentation.jpg" alt="Existing work: road-user segmentation on aerial footage"/><div class="caption">Existing work: road-user segmentation on aerial footage</div></div><div><h2>Perception comes before safety analysis</h2><p>A pretrained BEV model could reduce part of the manual perception effort. Tracking, calibration and conflict analysis still need their own validation.</p><div class="note" style="margin-top:24px">The long-term motivation remains road safety; the immediate contribution is a local benchmark.</div></div></div>
<div class="source">Background: <a href="https://github.com/ADLab-AutoDrive/BEVHeight">BEVHeight — official roadside 3D detection implementation</a>.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Research background</div>
<h1>What does “BEV” mean?</h1>
<div class="cols"><div><h2>Bird’s-eye view</h2><p>A common top-down coordinate system for describing roads, vehicle positions and movement.</p><ul><li>Input: camera images and calibration</li><li>Thesis output: 3D vehicle boxes and world-coordinate locations</li><li>Over time: positions can support tracking</li></ul><div class="note">A drone photo is an observation from above. A BEV model estimates a structured representation from its camera inputs.</div></div><div><img src="/assets/research-overview/bev-concept.svg" alt="Illustrative BEV output; not a prediction from this thesis"/><div class="caption">Illustrative BEV output; not a prediction from this thesis</div></div></div>
<div class="source">Background: <a href="https://github.com/ADLab-AutoDrive/BEVHeight">BEVHeight — official roadside 3D detection implementation</a>.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Research question</div>
<h1>What exactly will the benchmark answer?</h1>
<div class="cols"><div><h2>Proposed research question</h2><p style="font-size:36px">How accurately does BEVHeight detect vehicles in 3D and locate them in Dhaka road coordinates?</p><ul><li>Compare on the same held-out scenes</li><li>Report where each model succeeds or fails</li><li>Measure accuracy and computation cost</li></ul></div><div><div class="card"><h2>Primary model: BEVHeight</h2><p>A vision-based roadside 3D detector. Evaluate local vehicle classes, occlusion, camera geometry, day/night and fog.</p></div><div style="height:24px"></div><div class="card"><h2>Scope still to finalize</h2><p>Additional comparison models, class mapping, annotation protocol and acceptance criteria remain to be finalized.</p></div></div></div><div class="note">A model ranking and documented limitations are useful outcomes—even if no model is reliable in every condition.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · BEVHeight</div>
<h1>What is BEVHeight?</h1>
<div class="flow"><div class="card"><h2>Roadside image</h2><p>Observe vehicles from an elevated camera.</p></div><div class="arrow">→</div><div class="card"><h2>Height estimation</h2><p>Learn image features and height above the ground.</p></div><div class="arrow">→</div><div class="card"><h2>BEV representation</h2><p>Use calibrated geometry to place features in a top-down grid.</p></div><div class="arrow">→</div><div class="card"><h2>3D detection</h2><p>Predict vehicle boxes, classes and confidence scores.</p></div></div><div class="cols"><div><h2>Why it fits this thesis</h2><p>BEVHeight is designed for roadside 3D object detection. Its height-based view transformation is relevant to elevated CCTV viewpoints.</p><div class="note" style="margin-top:24px">The benchmark asks whether a pretrained model transfers reliably to Dhaka scenes and local vehicle classes.</div></div><div><img src="/assets/research-overview/bev-concept.svg" alt="Illustrative top-down road representation"/><div class="caption">Concept illustration, not BEVHeight predictions</div></div></div>
<div class="source">Source: <a href="https://openaccess.thecvf.com/content/CVPR2023/html/Yang_BEVHeight_A_Robust_Framework_for_Vision-Based_Roadside_3D_Object_Detection_CVPR_2023_paper.html">Yang et al., BEVHeight, CVPR 2023</a>.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · BEVHeight inputs</div>
<h1>What inputs does BEVHeight need?</h1>
<table><thead><tr><th>Input</th><th>Plain-language meaning</th><th>Source in this thesis</th></tr></thead><tbody><tr><td>RGB image</td><td>A roadside camera frame</td><td>CCTV or a rendered virtual CCTV view</td></tr><tr><td>Camera intrinsics</td><td>Focal lengths and image centre</td><td>Camera calibration or Blender camera settings</td></tr><tr><td>Camera pose</td><td>Camera orientation and location in the reference frame</td><td>Measured/calibrated pose or virtual camera transform</td></tr><tr><td>Ground reference</td><td>Ground alignment and reference camera height</td><td>Calibrated road plane or known scene geometry</td></tr><tr><td>Image transforms</td><td>How resizing, cropping or other preprocessing changed the image</td><td>Dataset preparation pipeline</td></tr></tbody></table><div class="note">LiDAR is not required at inference. The adapter must convert Blender axes, units and camera conventions to those expected by BEVHeight.</div>
<div class="source">Implementation: <a href="https://github.com/ADLab-AutoDrive/BEVHeight/blob/main/layers/backbones/lss_fpn.py">BEVHeight height-to-location geometry</a> · <a href="https://github.com/ADLab-AutoDrive/BEVHeight/blob/main/dataset/nusc_mv_det_dataset.py">Dataset inputs and calibration</a>.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Camera calibration</div>
<h1>Intrinsics: how the camera forms an image</h1>
<div class="cols"><div><img src="/assets/research-overview/camera-intrinsics.svg" alt="Image-plane diagram showing principal point coordinates and a projected pixel"/></div><div><h2>Parameters inside the camera model</h2><ul><li>fₓ, fᵧ: focal lengths in pixels</li><li>cₓ, cᵧ: principal point</li><li>Lens distortion: bending introduced by the lens</li></ul><div class="card" style="margin-top:24px"><p style="font-family:monospace;white-space:pre;line-height:1.6">K = [ fₓ 0 cₓ ]<br> [ 0 fᵧ cᵧ ]<br> [ 0 0 1 ]</p></div></div></div><div class="note">Resizing changes pixel-based intrinsics; cropping shifts the principal point. Keep calibration consistent with the image supplied to BEVHeight.</div><div class="source">Geometry: <a href="https://docs.opencv.org/4.13.0/d9/d0c/group__calib3d.html">OpenCV — Camera Calibration and 3D Reconstruction</a>. Diagrams are illustrative.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Camera calibration</div>
<h1>Extrinsics: where the camera is and where it points</h1>
<div class="cols"><div><img src="/assets/research-overview/camera-extrinsics.svg" alt="Roadside camera pose, viewing direction, height and world coordinate axes"/></div><div><h2>Transform between coordinate frames</h2><ul><li>R: rotation from world to camera axes</li><li>t: translation in that transformation</li></ul><div class="card" style="margin-top:24px"><p style="font-size:36px">X<sub>camera</sub> = R X<sub>world</sub> + t</p><p style="margin-top:18px">Camera centre in world coordinates:<br>C = −Rᵀt</p></div></div></div><div class="note">Moving or tilting a camera changes its extrinsics. Translation t is not the camera’s world position.</div><div class="source">Geometry: <a href="https://docs.opencv.org/4.13.0/d9/d0c/group__calib3d.html">OpenCV — Camera Calibration and 3D Reconstruction</a>. Diagrams are illustrative.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Camera calibration</div>
<h1>Calibration connects Blender, CCTV and BEVHeight</h1>
<div class="flow"><div class="card"><h2>World point</h2><p>Vehicle location in metres</p></div><div class="arrow">→ R, t →</div><div class="card"><h2>Camera point</h2><p>Location relative to the camera</p></div><div class="arrow">→ K →</div><div class="card"><h2>Image pixel</h2><p>Projected location after perspective division</p></div></div><div class="cols"><div class="card"><h2>Real CCTV — proposed preparation</h2><p>Estimate lens calibration and pose against measured road references. Check projected landmarks on the image.</p></div><div class="card"><h2>Blender — proposed export</h2><p>Derive intrinsics from camera/render settings. Convert camera transforms to the model’s axes and units; export calibration with every view.</p></div></div><div class="note">A pixel defines a viewing ray, not a unique 3D location. BEVHeight supplies learned height information; calibrated geometry connects it to the road frame.</div><div class="source">Geometry: <a href="https://docs.opencv.org/4.13.0/d9/d0c/group__calib3d.html">OpenCV — Camera Calibration and 3D Reconstruction</a>. Diagrams are illustrative.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · World locations</div>
<h1>From a camera image to a location on the road</h1>
<div class="flow">
<div class="card"><h2>CCTV image</h2><p>Vehicle appears in perspective; pixels alone are not metres.</p></div><div class="arrow">→</div>
<div class="card"><h2>BEVHeight</h2><p>Predict a 3D vehicle box in the model’s coordinate convention.</p></div><div class="arrow">→</div>
<div class="card"><h2>Calibrated transform</h2><p>Convert the predicted location to one shared road coordinate frame.</p></div><div class="arrow">→</div>
<div class="card"><h2>World location</h2><p>Vehicle position in metres, ready to compare with the reference.</p></div>
</div>
<div class="cols"><div><img src="/assets/research-overview/bev-concept.svg" alt="Schematic road coordinate frame and vehicle paths"/></div><div><h2>Use one location definition</h2><p>For example, the ground projection of the 3D box centre. Use the same point for predictions and reference labels.</p><div class="note" style="margin-top:24px">Local road coordinates are sufficient for evaluation. Latitude/longitude requires additional georeferencing.</div></div></div>
<div class="source">Model: <a href="https://openaccess.thecvf.com/content/CVPR2023/html/Yang_BEVHeight_A_Robust_Framework_for_Vision-Based_Roadside_3D_Object_Detection_CVPR_2023_paper.html">Yang et al., BEVHeight, CVPR 2023</a>. Coordinate alignment is part of the proposed evaluation.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Inputs and labels</div>
<h1>Model inputs and reference labels have different roles</h1>
<div class="cols"><div class="card"><h2>Prediction branch</h2><p>RGB image + calibration<br>↓<br>Pretrained BEVHeight<br>↓<br>Predicted 3D box and confidence<br>↓<br>Location in the shared road frame</p></div><div class="card"><h2>Reference branch</h2><p>Known simulated vehicle geometry and pose<br>or reviewed real-data annotations<br>↓<br>Reference boxes and locations<br>↓<br>Compare with the model’s predictions</p></div></div><div class="note">Labels are targets during supervised training and references during evaluation. They are not supplied to the model as answers at inference.</div><p>Nadir imagery supports real-world reference construction; the primary detection input is the roadside view. Map Dhaka vehicle categories to supported model classes explicitly.</p>
<div class="source">Implementation: <a href="https://github.com/ADLab-AutoDrive/BEVHeight/blob/main/layers/backbones/lss_fpn.py">BEVHeight height-to-location geometry</a> · <a href="https://github.com/ADLab-AutoDrive/BEVHeight/blob/main/dataset/nusc_mv_det_dataset.py">Dataset inputs and calibration</a>.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Reconstruction method</div>
<h1>Why reconstruct roads without LiDAR?</h1>
<div class="cols"><div><img src="/assets/week-11/daytime-no-fog.png" alt="Own Blender point-cloud scene of Polashi Intersection"/><div class="caption">Own Blender point-cloud scene of Polashi Intersection</div></div><div><h2>Available resource: camera imagery</h2><p>Overlapping drone images can recover camera poses and scene structure. A reconstructed local scene can support repeatable viewpoint experiments.</p><ul><li>Capture local geometry and appearance</li><li>Reuse the scene for controlled rendering</li><li>Investigate camera placement and visibility</li></ul><div class="note">Metric distances need scale control, such as measured reference distances. Reconstruction accuracy must be checked independently.</div></div></div>
<div class="source">Method background: <a href="https://colmap.github.io/tutorial.html">COLMAP reconstruction tutorial</a>.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Reconstruction method</div>
<h1>Where COLMAP and Gaussian splatting fit</h1>
<div class="flow"><div class="card"><h2>1 · Capture</h2><p>Overlapping images of an intersection or road</p></div><div class="arrow">→</div><div class="card"><h2>2 · COLMAP</h2><p>Estimate camera poses and 3D points</p></div><div class="arrow">→</div><div class="card"><h2>3 · 3D Gaussian splatting</h2><p>Optimize a scene representation from images and poses</p></div><div class="arrow">→</div><div class="card"><h2>4 · Render</h2><p>Generate views for controlled experiments</p></div></div><div class="cols"><div><img src="/assets/image7.png" alt="Earlier reconstruction: point cloud before the color fix"/><div class="caption">Earlier reconstruction: point cloud before the color fix</div></div><div><h2>Current evidence and next milestone</h2><p>Current slides demonstrate a colored Blender point cloud and scene edits. Record the actual Gaussian-splat training and rendering outputs as a separate milestone.</p><div class="note" style="margin-top:24px">COLMAP reconstructs geometry. Gaussian splatting adds learned appearance for novel views. Neither automatically creates BEV labels or moving traffic.</div></div></div>
<div class="source">Sources: <a href="https://colmap.github.io/tutorial.html">COLMAP</a> · <a href="https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/">Kerbl et al., 3D Gaussian Splatting, 2023</a>.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Evaluation plan</div>
<h1>Two complementary evaluation routes</h1>
<div class="cols"><div class="card"><h2>A · Real Dhaka data</h2><p>Synchronized CCTV + nadir drone views<br>↓<br>Calibrated and reviewed reference labels<br>↓<br>Model predictions<br>↓<br>Errors on unseen local scenes</p><div class="note" style="margin-top:24px">Real-world cross-check of detection and location</div></div><div class="card"><h2>B · Reconstructed scenes</h2><p>Gaussian-splat background + labeled vehicles<br>↓<br>Known virtual cameras and object transforms<br>↓<br>Model predictions<br>↓<br>Stress-test sensitivity</p><div class="note" style="margin-top:24px">Exact simulation labels; real-world validity still tested</div></div></div><p>Split by intersection or capture session before reconstruction or model tuning. Nearby frames from the same scene must not leak across train and test sets.</p>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Ground truth</div>
<h1>How the proposed ground truth is constructed</h1>
<div class="cols"><div class="card"><h2>Simulation: known 3D labels</h2><ul><li>Scale and align the reconstructed road.</li><li>Add vehicles with known dimensions, pose and object IDs.</li><li>Render CCTV-like views with known camera calibration.</li><li>Export boxes and world positions directly from scene transforms.</li></ul><div class="note">The scene’s object metadata supplies labels. Gaussian-splat appearance alone does not.</div></div><div class="card"><h2>Real footage: nadir + CCTV</h2><ul><li>Synchronize captures of the same traffic event.</li><li>Calibrate the nadir view to measured road coordinates.</li><li>Match and annotate vehicles across both views.</li><li>Compare CCTV predictions with the aligned reference positions.</li></ul><div class="note">Nadir views support planar position labels. Full 3D boxes need dimensions, height and orientation evidence.</div></div></div>
<p>Keep simulation scores and real-data scores separate. Report reference uncertainty and exclude unmatched or unobservable cases using a documented rule.</p>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Benchmark protocol</div>
<h1>Make the comparison fair and measurable</h1>
<table><thead><tr><th>Decision</th><th>Proposed evaluation rule</th></tr></thead><tbody><tr><td>Camera input</td><td>Match each model’s required camera count, calibration and viewpoint.</td></tr><tr><td>Primary task</td><td>3D vehicle detection and world-location extraction; tracking is downstream.</td></tr><tr><td>Reference labels</td><td>Use independent annotations and measured scale; document uncertainty.</td></tr><tr><td>Metrics</td><td>3D/BEV average precision; matched-vehicle location error in metres; recall.</td></tr><tr><td>Conditions and cost</td><td>Report results by scene/condition, plus runtime on the same hardware.</td></tr></tbody></table><div class="note">BEVHeight targets roadside cameras. Match its calibration, ground-plane convention and class definitions; use nadir views as a reference source.</div>
<div class="source">Background: <a href="https://github.com/ADLab-AutoDrive/BEVHeight">BEVHeight — official roadside 3D detection implementation</a>. Evaluation rules shown here are a proposed thesis protocol.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Synthetic dataset</div>
<h1>Blender as a synthetic dataset generator</h1>
<div class="flow"><div class="card"><h2>Build the scene</h2><p>Scaled local road background + separate vehicle assets</p></div><div class="arrow">→</div><div class="card"><h2>Set the scenario</h2><p>Choose cameras, vehicle poses, lighting and fog</p></div><div class="arrow">→</div><div class="card"><h2>Render and export</h2><p>Save RGB views, camera data and object labels</p></div><div class="arrow">→</div><div class="card"><h2>Check and package</h2><p>Validate labels and convert to the BEVHeight dataset format</p></div></div><div class="cols"><div><img src="/assets/week-11/daytime-no-fog.png" alt="Existing Blender scene used as the foundation for the proposed synthetic dataset"/><div class="caption">Existing scene foundation; automatic dataset export is proposed</div></div><div><h2>One scene, paired observations</h2><p>Render a roadside view and a nadir view of the same simulation state. Both share the same vehicle IDs and world coordinates.</p><div class="note" style="margin-top:24px">Use explicit vehicle objects for labels. Remove or mask vehicles already captured in the static splat background.</div></div></div>
<div class="source">Blender capabilities: <a href="https://docs.blender.org/manual/en/latest/render/layers/passes.html">Render passes and object masks</a>. Dataset workflow and variation counts below are proposed for this thesis.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Dataset variations</div>
<h1>Generate many variations systematically</h1>
<table><thead><tr><th>Variation</th><th>Examples to control</th><th>What the benchmark can test</th></tr></thead><tbody><tr><td>Camera</td><td>Height, tilt, position and focal length</td><td>Sensitivity to viewpoint and calibration</td></tr><tr><td>Vehicles</td><td>Type, dimensions, orientation, count and placement</td><td>Class coverage, distance and occlusion</td></tr><tr><td>Lighting</td><td>HDRI, sun direction and night lights</td><td>Performance under appearance changes</td></tr><tr><td>Visibility</td><td>No fog, low fog and high fog</td><td>Detection loss as visibility declines</td></tr><tr><td>Image quality</td><td>Exposure, blur and compression</td><td>Sensitivity to CCTV image degradation</td></tr></tbody></table><div class="note">Illustrative plan: 5 camera poses × 3 lighting settings × 3 fog levels × 20 traffic layouts = 900 roadside images per scene.</div><p class="caption">Save random seeds and settings. Keep all variants of a scene in the same dataset split; 900 renders are not 900 independent real scenes.</p>
<div class="source">Blender capabilities: <a href="https://docs.blender.org/manual/en/latest/render/layers/passes.html">Render passes and object masks</a>. Dataset workflow and variation counts below are proposed for this thesis.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Dataset validation</div>
<h1>Every rendered sample needs matching labels</h1>
<div class="cols"><div><h2>Export a complete sample</h2><ul><li>RGB frame, scene ID and scenario settings</li><li>Camera intrinsics, pose and ground reference</li><li>Vehicle ID, class, 3D dimensions and orientation</li><li>World position using a consistent anchor point</li><li>Optional depth and object masks for visibility checks</li></ul></div><div><h2>Validate before benchmarking</h2><ul><li>Project 3D box corners onto the RGB image.</li><li>Check metres, axes and camera alignment.</li><li>Flag truncation and heavy occlusion.</li><li>Keep held-out real CCTV/nadir data for transfer testing.</li></ul></div></div><div class="note">Ordinary Gaussian splats contain captured appearance. Physically consistent relighting, shadows and fog may require a compatible renderer or a hybrid scene with conventional geometry; verify the integration.</div><p>Start with a small, manually checked batch. Scale generation only after images, calibration and labels agree.</p>
<div class="source">Blender capabilities: <a href="https://docs.blender.org/manual/en/latest/render/layers/passes.html">Render passes and object masks</a>. Dataset workflow and variation counts below are proposed for this thesis. 3D boxes require a custom metadata exporter; RGB rendering alone does not produce them.</div>
</div>
---
class: week11-light
---
<!-- Week 11 — Color Fix -->
<div class="flex h-full flex-col justify-center gap-6 px-12">
<h1 class="text-6xl font-bold">Revisiting the Color Problem</h1>
<div class="text-3xl">Revisiting slide 8 from the earlier progress update: the reconstructed point cloud appeared without its original colors.</div>
<img class="mx-auto w-full object-contain" style="height: 680px" src="/assets/image7.png" alt="Original screenshot from slide 8 showing the reconstructed point cloud appearing gray in Blender" />
<div class="text-2xl">Week 11: restored color display with fix_ply_colors.py, then refined the materials and scene manually.</div>
</div>
---
class: week11-light
---
<!-- Week 11 — Color Fix and Scene Refinement -->
<div class="flex h-full flex-col justify-center gap-10 px-12">
<h1 class="text-6xl font-bold">Point Cloud Color Fix &amp; Scene Refinement</h1>
<div class="grid grid-cols-2 gap-16">
<div>
<h2 class="text-4xl font-bold">Initial script setup</h2>
<div class="mt-4 text-2xl">fix_ply_colors.py</div>
<ul class="mt-6 list-disc space-y-4 pl-8 text-3xl leading-relaxed">
<li>Converted imported PLY geometry to a Blender point cloud</li>
<li>Connected the Col color attribute to an emission material</li>
<li>Added adjustable point radius through Geometry Nodes</li>
<li>Configured the camera and saved the Blender scene</li>
</ul>
</div>
<div>
<h2 class="text-4xl font-bold">Manual changes after the script</h2>
<ul class="mt-6 list-disc space-y-4 pl-8 text-3xl leading-relaxed">
<li>Replaced the emission shader with Principled BSDF</li>
<li>Added a sky HDRI to the world environment</li>
<li>Added fog using Principled Volume</li>
<li>Removed some reconstruction artifacts</li>
</ul>
</div>
</div>
</div>
---
class: week11-light
---
<!-- Week 11 — Weather and Lighting Tests -->
<div class="flex h-full flex-col justify-center gap-10 px-12">
<h1 class="text-6xl font-bold">Preliminary Weather &amp; Lighting Tests</h1>
<ul class="list-disc space-y-4 pl-8 text-4xl leading-relaxed">
<li>Input: previously acquired photogrammetry data</li>
<li>Lighting conditions: day and night</li>
<li>Weather conditions: fog and no fog</li>
<li>Status: preliminary tests conducted</li>
</ul>
</div>
---
class: week11-light
---
<!-- Week 11 — Daytime Comparison -->
<div class="flex h-full flex-col justify-center gap-10 px-12">
<h1 class="text-6xl font-bold">Daytime — No Fog / Intense Fog</h1>
<div class="grid grid-cols-2 gap-8">
<div class="flex flex-col gap-5">
<div class="text-3xl font-bold">No fog</div>
<img class="w-full object-contain" style="height: 600px" src="/assets/week-11/daytime-no-fog.png" alt="Blender editor screenshot of the daytime point cloud without fog" />
</div>
<div class="flex flex-col gap-5">
<div class="text-3xl font-bold">Intense fog</div>
<img class="w-full object-contain" style="height: 600px" src="/assets/week-11/daytime-intense-fog.png" alt="Blender editor screenshot of the daytime point cloud with intense fog" />
</div>
</div>
<div class="text-2xl text-slate-300">Preliminary Blender tests using previously acquired photogrammetry data</div>
</div>
---
class: week11-light
---
<!-- Week 11 — Night Comparison -->
<div class="flex h-full flex-col justify-center gap-10 px-12">
<h1 class="text-6xl font-bold">Night — Fog Comparison</h1>
<div class="grid grid-cols-3 gap-6">
<div class="flex flex-col gap-5">
<div class="text-3xl font-bold">No fog</div>
<img class="w-full object-contain" style="height: 420px" src="/assets/week-11/night-no-fog.png" alt="Night render of the reconstructed scene without fog" />
</div>
<div class="flex flex-col gap-5">
<div class="text-3xl font-bold">Low fog</div>
<img class="w-full object-contain" style="height: 420px" src="/assets/week-11/night-low-fog.png" alt="Night render of the reconstructed scene with low fog" />
</div>
<div class="flex flex-col gap-5">
<div class="text-3xl font-bold">High fog</div>
<img class="w-full object-contain" style="height: 420px" src="/assets/week-11/night-high-fog.png" alt="Night render of the reconstructed scene with high fog" />
</div>
</div>
<div class="text-2xl text-slate-300">Preliminary renders under simulated night lighting and fog</div>
</div>
---
class: week11-light
---
<!-- Week 11 — Real Fog Reference -->
<div class="flex h-full flex-col justify-center gap-10 px-12">
<h1 class="text-6xl font-bold">Daytime Fog — Visual Reference</h1>
<div class="grid grid-cols-2 gap-8">
<div class="flex flex-col gap-5">
<div class="text-3xl font-bold">Blender — intense fog</div>
<img class="w-full object-contain" style="height: 600px" src="/assets/week-11/daytime-intense-fog.png" alt="Simulated intense daytime fog in the Blender editor" />
</div>
<div class="flex flex-col gap-5">
<div class="text-3xl font-bold">Real photograph — intense fog</div>
<img class="w-full object-contain" style="height: 600px" src="/assets/week-11/real-daytime-fog.webp" alt="Real fog reference photograph showing an auto-rickshaw and a cyclist in reduced visibility" />
</div>
</div>
<div class="text-2xl text-slate-300">Qualitative appearance reference; different scenes and viewpoints</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Interpreting the tests</div>
<h1>What the weather experiments contribute</h1>
<div style="display:grid;grid-template-columns:repeat(3,1fr);gap:24px;margin-top:35px"><div><img src="/assets/week-11/night-no-fog.png" alt="Own Blender render · night / no fog"/><div class="caption">Own Blender render · night / no fog</div></div><div><img src="/assets/week-11/night-low-fog.png" alt="Own Blender render · night / low fog"/><div class="caption">Own Blender render · night / low fog</div></div><div><img src="/assets/week-11/night-high-fog.png" alt="Own Blender render · night / high fog"/><div class="caption">Own Blender render · night / high fog</div></div></div><div class="cols"><div class="card"><h2>Question to test</h2><p>How much does model accuracy change as visibility worsens in the same scene?</p></div><div class="card"><h2>What still needs validation</h2><p>Document fog and lighting settings, and keep labeled vehicle positions fixed. These images demonstrate scene setup; detection accuracy is not yet measured.</p></div></div><div class="note">The splat supplies the road background. Add independently labeled 3D vehicles to evaluate detection and world location; avoid unlabeled vehicles baked into the background.</div>
</div>
---
class: week11-light
---
<!-- Week 11 — Drone Capture DJI 0403 -->
<div class="flex h-full flex-col justify-center gap-6 px-12">
<h1 class="text-6xl font-bold">New Drone Capture — DJI_0403</h1>
<div class="text-3xl">Newly captured footage, stabilized for trajectory analysis</div>
<video class="w-full object-contain" style="height: 760px" src="/assets/week-11/dji-0403-stabilized-2k.mp4" controls autoplay loop muted playsinline preload="metadata" aria-label="DJI 0403 drone capture and stabilization preview"></video>
<div class="text-2xl">Stabilization preview · First 20 seconds</div>
</div>
---
class: week11-light
---
<!-- Week 11 — Drone Capture DJI 0406 -->
<div class="flex h-full flex-col justify-center gap-6 px-12">
<h1 class="text-6xl font-bold">New Drone Capture — DJI_0406</h1>
<div class="text-3xl">Newly captured footage, stabilized for trajectory analysis</div>
<video class="w-full object-contain" style="height: 760px" src="/assets/week-11/dji-0406-stabilized-2k.mp4" controls autoplay loop muted playsinline preload="metadata" aria-label="DJI 0406 drone capture and stabilization preview"></video>
<div class="text-2xl">Stabilization preview · First 20 seconds</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Future safety application</div>
<h1>After validation: return to safety analysis</h1>
<div class="flow"><div class="card"><h2>BEV perception</h2><p>Road-user locations in a metric frame</p></div><div class="arrow">→</div><div class="card"><h2>Tracking</h2><p>Paths, speed and timing</p></div><div class="arrow">→</div><div class="card"><h2>Conflict measures</h2><p>Time-to-collision / post-encroachment time</p></div><div class="arrow">→</div><div class="card"><h2>Engineering review</h2><p>Find recurring conflict locations</p></div></div><div class="cols"><div><img src="/assets/research-overview/bev-concept.svg" alt="Illustration of paths in a shared coordinate system"/><div class="caption">Illustration of paths in a shared coordinate system</div></div><div><h2>Example: crossing movements</h2><p>Estimate when two road users approach or pass through a shared conflict area. Review recurring interactions by location and time.</p><div class="note" style="margin-top:24px">Requires validated trajectories and timestamps. Conflict indicators are surrogate measures; they do not directly prove a reduction in crashes.</div></div></div>
<div class="source">Safety-method background: <a href="https://www.fhwa.dot.gov/publications/research/safety/08051/02.cfm">FHWA — Surrogate Safety Assessment Model, Chapter 2</a>.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Bangladesh applications</div>
<h1>Where this could be used in Bangladesh</h1>
<table><thead><tr><th>Potential setting / user</th><th>Possible use after validation</th><th>Decision supported</th></tr></thead><tbody><tr><td>Dhaka intersections · traffic police</td><td>Turning counts, queues and recurring conflicts</td><td>Review signal timing and traffic control</td></tr><tr><td>Urban streets · city corporations</td><td>Pedestrian–vehicle interaction maps</td><td>Prioritize crossing and junction audits</td></tr><tr><td>Highway junctions · RHD</td><td>Site-level movement and visibility studies</td><td>Prioritize locations for engineering inspection</td></tr><tr><td>BRTA / university research teams</td><td>Reusable local datasets and model comparisons</td><td>Choose models based on local evidence</td></tr><tr><td>Transport planning teams</td><td>Reconstructed scenes and observed traffic overlays</td><td>Communicate design options and test scenarios</td></tr></tbody></table><div class="note">These are proposed applications, not current deployments or agency partnerships.</div>
<div class="source">Bangladesh context: <a href="https://www.worldbank.org/en/news/press-release/2023/06/14/bangladesh-and-the-world-bank-joins-hands-to-improve-road-safety">World Bank / Bangladesh road-safety project: cities, high-risk highways and district roads</a>. Application mapping is a proposed extension of this thesis.</div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Digital twin potential</div>
<h1>A path toward a traffic digital twin</h1>
<div class="flow"><div class="card"><h2>Scene foundation</h2><p>Reconstructed road geometry and appearance</p></div><div class="arrow">→</div><div class="card"><h2>Traffic layer</h2><p>Validated positions and trajectories</p></div><div class="arrow">→</div><div class="card"><h2>Updated system</h2><p>Connect observations over time</p></div><div class="arrow">→</div><div class="card"><h2>Planning tools</h2><p>Add and validate scenario simulation</p></div></div><div class="cols"><div><img src="/assets/week-11/daytime-no-fog.png" alt="Current scene foundation — Blender reconstruction"/><div class="caption">Current scene foundation — Blender reconstruction</div></div><div><h2>Example future use</h2><p>Overlay observed turning movements on a local intersection, then study a proposed crossing or signal plan using a validated traffic simulation.</p><div class="note" style="margin-top:24px">The current 3D scene is a foundation. A live digital twin also needs updated data, traffic behavior and validation of its predictions.</div></div></div>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Thesis deliverables</div>
<h1>What a successful thesis will deliver</h1>
<div class="cols"><div><h2>Core research deliverables</h2><ul><li>A documented Dhaka evaluation dataset and labeling protocol</li><li>A reproducible comparison of selected compatible BEV models</li><li>Results by scene and condition, with failure examples</li><li>Evidence on the usefulness and limits of reconstructed scenes</li></ul></div><div><h2>Current work → next evidence</h2><div class="card"><h2>Completed groundwork</h2><p>Drone capture and stabilization; point-cloud color fix; manual materials, HDRI, fog and artifact cleanup.</p></div><div style="height:20px"></div><div class="card"><h2>Next</h2><p>Set up BEVHeight; align CCTV and nadir references; add labeled simulation vehicles; verify reconstruction; run the held-out benchmark.</p></div></div></div><div class="note">My thesis asks which BEV models are reliable for Dhaka. The reconstruction is the experimental support; safety analysis and digital twins are downstream opportunities.</div>
</div>
---
class: week11-light
---
<!-- Week 11 — Next Steps -->
<div class="flex h-full flex-col justify-center gap-10 px-12">
<h1 class="text-6xl font-bold">Next Steps</h1>
<ol class="list-decimal space-y-4 pl-12 text-4xl leading-relaxed">
<li>Recapture imagery and improve the COLMAP reconstruction</li>
<li>Verify Gaussian-splat rendering and labeled Blender scenarios</li>
<li>Calibrate and align CCTV and nadir reference data</li>
<li>Prepare BEVHeight and run the held-out detection benchmark</li>
</ol>
</div>
---
class: week11-light
---
<div class="research-slide">
<div class="eyebrow">Week 11 progress · Discussion</div>
<h1>Thank You</h1>
<div style="flex:1;display:flex;flex-direction:column;justify-content:center;gap:40px"><p style="font-size:52px;font-weight:800;max-width:1450px">Week 11 Progress — Discussion</p><p>BEVHeight baseline · simulation labels · CCTV–nadir alignment</p></div>
</div>

Xet Storage Details

Size:
48.8 kB
·
Xet hash:
5f4145837305cca58baa282e476471ff5e7c7e63decd8eb43e37e1aad7c3adf4

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.