With the Quarterly Review Board (QRB) presentation coming up, the team is fully focused on polishing our 3D modeling pipeline and ensuring our demo is ready for the judges.
Boosting Model Quality We made a key decision to utilize the full image dataset for our reconstructions. We confirmed that using the complete set of frames fills in critical gaps and captures much finer details—like the texture of the statue—without adding significant processing time. The result is a much smoother and more complete 3D model.
Focus on Cleanup & Integration Our current priority is eliminating background noise to ensure a clean final look. We are testing new methods to automatically remove these “outliers” and are finalizing the texture mapping process to get everything ready for integration. We are excited to showcase our progress to the board in two weeks!
This week was a major turning point for the IPPD VOXEL team. During our update meeting, we revealed our latest 3D reconstruction results using the “Ignatius” model, and the reaction was incredible. The rendering quality has improved so drastically that our coach, Catia Silva, explicitly mentioned getting “goosebumps” from the smooth details—even the folds on the statue’s collar were clearly visible.
What makes this even more exciting is that we achieved this high level of fidelity without any complex manual tuning yet, proving that our core pipeline is solid. Beyond the technical win, our team was also praised for our strong collaboration. Unlike other groups where workload issues were visible, our coach noted that our team clearly shares ownership and works seamlessly together.
Preparations for the upcoming Quarterly Review Board (QRB) are also proceeding smoothly. Since our last presentation, we have successfully integrated the system and significantly improved the 3D mesh quality, so we are confident in demonstrating our substantial progress to the judges. We are now more confident than ever that we will deliver a top-notch product by March!
This week focused on reviewing our progress during the first project milestone through QRB1, which provided an opportunity to validate our direction and overall approach. The presentation was well received and clearly communicated our objectives, progress, and planned next steps, reinforcing confidence in the current project structure.
Following the review, we shifted our attention toward exploring 2D Gaussian Splatting as a potential approach to address ongoing challenges related to surface quality and mesh consistency. This exploration was done in the context of the existing pipeline, with consideration given to how it could complement segmentation and 3D reconstruction stages.
Overall, this week served as a transition from planning and alignment into deeper technical exploration, helping set the foundation for focused experimentation in the coming weeks.
This week, our team revisited the mesh generation stage of the project and discussed the challenges involved in converting 3D reconstructions into usable assets. Based on these discussions, we explored how different components of the pipeline, such as segmentation and 3D reconstruction, influence mesh quality and overall asset usability.
To improve efficiency, we reorganized our workload by splitting responsibilities across team members and prioritizing tasks that have the highest impact on the project outcomes. This allowed us to focus more attention on high-priority areas while ensuring parallel progress across supporting tasks.
These planning decisions helped align our technical direction with project goals and improved clarity on next steps as we continue refining the pipeline
Week 1 of IPPD 2 felt a bit like reopening a familiar book after a long break. As we gathered in the room again, there was that mix of catching up as friends and reconnecting as teammates. We talked about where we left off last semester, our early prototypes, the milestones we hit, and the parts that were tougher than expected. Remembering the late nights, quick fixes, and last-minute rehearsals helped us see how far we’ve already come as a team, not just in terms of the project, but in how we communicate and work together.
From there, the conversation naturally shifted toward what comes next. We started sketching out our goals for this semester: what done should look like, which pieces of the system still need to be built or polished, and how we’re going to divide responsibilities so no one feels overloaded. Instead of just listing tasks, we tried to think in terms of phases and stories what we want to be able to show at the next review, what would make our sponsor excited, and how we can turn last semester’s lessons into smarter decisions this time. By the end of the meeting, it felt like we had turned the page and begun a new chapter, with a clearer sense of direction and renewed energy for the work ahead.
This week felt like the “finale chapter” of our VOXEL O2A journey, centered around our final System-Level Design Review (SLDR). The afternoon began with a networking session from 1–3 PM, where a liaison sat at our table and spoke candidly about what life looks like in the real world beyond the classroom. Instead of talking about our specific project, he focused on broader themes: how expectations change when you enter industry, the importance of reliability and consistency, and why skills like communication, accountability, and learning quickly on the job matter just as much as technical knowledge. Hearing about challenges such as handling ambiguity, working with diverse teams, and managing deadlines gave us a clearer picture of what it really means to be ready to work, not just technically strong.
From 3–5 PM, we shifted into our final SLDR presentation. As a team, we walked the audience through our full pipeline story – starting from a simple phone capture and moving through segmentation, 3D reconstruction, mesh and texture generation, and finally into Unity, where assets can be viewed, rotated, and inspected and also we had the exciting opportunity to demo our working prototype, showcasing our end-to-end pipeline that transforms standard video footage into 3D meshes using Gaussian Splatting. It was incredibly rewarding to see the pipeline in action successfully processing the “Gator” video from segmentation to mesh extraction and to receive validation that our progress is well ahead of the curve for this stage of the project.
Beyond the technical success, the highlight of the day was undoubtedly the human connection. We finally had the chance to meet our liaisons, in person after weeks of remote collaboration, which added a great dynamic to our discussions about future use cases and optimizations. The event was a fantastic experience overall. it was fun networking with new people, sharing our work with a broader audience, and celebrating the solid foundation we’ve built as we head into the next phase of development.
This week was all about our PR System-Level Design Review (SLDR) for the VOXEL O2A pipeline. We presented alongside four other IPPD teams, our coaches, and the TA, so it felt more like a mini conference than a regular class review. As a team, we walked everyone through how a simple phone capture travels through our pipeline from object isolation and 3D reconstruction all the way to Unity-ready assets. People responded well to the way we framed O2A as a complete, mobile first workflow rather than a single demo, and several reviewers specifically highlighted our clear module breakdown, error-handling strategy, and focus on turning research ideas into something a real user could actually touch and use.
The overall tone of the feedback was very encouraging. Reviewers felt we had a solid foundation and a realistic plan for where the system is headed next. The main suggestions were to simplify our architecture slide so that the core data flow pops out more clearly at a glance, and to spell out our testing plan in a more structured way what datasets we’ll use, which metrics (like PSNR, SSIM, runtime) we’ll report, and how we’ll compare against baselines such as NeRF and COLMAP. Both of these feel more like polishing than major rework, which boosted our confidence that we’re on the right track.
Looking ahead, we want to convert this momentum into concrete upgrades. We plan to redesign the architecture slide with cleaner visuals, then expand our SLDR document to clearly describe our evaluation pipeline, baselines, and Spring 2026 testing roadmap. In parallel, we’ll keep strengthening the technical side: stabilizing our mesh and texture pipeline, experimenting with Poisson/Ball Pivoting, running 3D Gaussian Splatting on benchmark datasets, and tightening our Unity-based visualization so people can freely rotate and inspect our reconstructed assets. Overall, this SLDR made VOXEL O2A feel less like a one-time prototype and more like a product-ready pipeline in progress, and that’s a direction we’re excited to keep building toward.
This week, our main focus was on mesh generation and understanding how to turn point-based or radiance field outputs into clean, usable 3D assets. We experimented with multiple mesh extraction methods, and spent time tuning parameters like density, smoothing, and decimation to see how they affect both visual quality and downstream usability. Instead of treating mesh as a “black box” step at the end of the pipeline, we dug into how surface reconstruction actually works and what kinds of inputs (noise level, point distribution, normal estimates) lead to stable outputs. This deeper understanding is already helping us recognize why some scenes reconstruct well while others collapse into artifacts or holes.
On the integration side, we started wiring these mesh outputs back into our VOXEL workflow so that meshes are not just generated in isolation but produced as a repeatable stage in the pipeline. We compared meshes across different methods for the same scene, checked how well they preserve fine details, and considered how they will behave once textured and visualized in tools like Unity. As we refine this stage, we can clearly see the full pipeline coming together: capture → segmentation → reconstruction → mesh → (soon) texturing and interactive viewing. At this point, we feel we are only two key steps away from an initial end-to-end pipeline by the end of the semester, and this week’s mesh work was a critical bridge between our earlier 3DGS experiments and a product-ready asset workflow.
This week was all about Prototype Inspection Day (PID). Our team presented the VOXEL O2A pipeline to six judges and walked them through how we turn phone captured images into 3D assets. Two of us presented online while the rest of the team was in person, so it ended up being a mix of virtual and live interaction, but everything ran smoothly. The overall response was very positive. A few judges even mentioned that the project felt 99% there, which made us feel confident that our idea and pipeline are on the right track.
At the same time, the feedback clearly showed where we can improve. The judges wanted to interact with our results instead of only seeing static screenshots. They asked to rotate the 3D models, view them from different angles, and understand what the final user experience would look like. They also suggested that we show clearer comparisons against baseline methods like NeRF and COLMAP, include metrics such as PSNR, SSIM and runtime, and add more context about our segmentation and evaluation setup. We also realized we need to balance our presentation style. Some people felt parts were too technical, while others wanted more detail, so we plan to layer our explanations for different audiences.
Looking ahead to next week, we want to turn this feedback into concrete progress. On the technical side, we will keep moving toward a stable end to end mesh and texture pipeline, test mesh extraction methods like Poisson and Ball Pivoting in Meshlab, and see which ones give cleaner assets. We will also start running 3D Gaussian Splatting on the NeRF synthetic dataset and explore Unity based visualization so people can view and rotate our reconstructed meshes directly and compare quality and runtime. In parallel, we will begin drafting the SLDR and make sure it clearly explains our testing strategy, roadmap and baselines so VOXEL O2A feels less like a demo and more like a product ready workflow.
team VOXEL presenting our 3D reconstruction framework at UF AI Days 2025
This week, our team successfully finalized and presented our poster at the UF AI Days Showcase, where we summarized our comparative study of NeRF variants. During the presentation, we highlighted the strengths and limitations of each model. For use cases requiring faster reconstruction, we recommended alternative NeRF variants optimized for speed while maintaining reasonable quality.
We are now successfully concluding Phase 1 of our project, which encompasses the end-to-end workflow from scene segmentation to NeRF-based 3D reconstruction. Testing and evaluation are actively in progress, and the initial results are highly promising. Our current system demonstrates strong structural fidelity and visual accuracy, consistently generating precise 3D reconstructions.
We have observed a notable improvement in both processing speed and model compactness compared to previously established methods. Specifically, the time required for both preparatory steps and core training has been significantly reduced. These strong outcomes strongly indicate that our core development, the adaptive densification and pruning strategy is highly effective. It successfully strikes a fine balance between computational efficiency and superior accuracy while ensuring the resulting model remains well optimized.
Overall, this week marks a major milestone as we transition from model development to full system integration and evaluation. In the next phase, we plan to extend testing to segmented, mobile-captured objects and explore alternative reconstruction techniques such as DUSt3R, aiming to further enhance both reconstruction quality and runtime performance.
