Jiawei Ren
@jiawei6_ren
Followers
1K
Following
531
Statuses
82
PhD student at @MMLabNTU
Singapore
Joined October 2021
RT @wilson_over: Say goodbye to perfect pinhole assumptions Excited to introduce 3DGUT—a Gaussian Splatting formulation that unlocks suppo…
0
42
0
🔥L4GM code and model weights are finally released! !🔥 Try it and turn your video into a 3D animation in just seconds! Code: Models:
We present #L4GM, the first 4D Large Reconstruction Model that produces animated objects from a single-view video input -- in a single feed-forward pass that takes only **seconds**! 1/
1
21
115
RT @janusch_patas: Feed-Forward Bullet-Time Reconstruction of Dynamic Scenes from Monocular Videos TL;DR: We present the first feed-forwar…
0
11
0
RT @HuanLing6: Checkout our latest 4D reconstrucion paper BTimer! Amazing work from @jiawei6_ren @huangjh_hjh and the team!
0
1
0
RT @WilliamLamkin: BTimer: Feed-Forward Bullet-Time Reconstruction of Dynamic Scenes from Monocular Videos website:
0
2
0
RT @hx_liang95: 🚀Excited to Introduce #BTimer: Real-Time Dynamic Scene Reconstruction from Monocular Videos! Struggling with novel view…
0
16
0
RT @RadianceFields: New research from @NVIDIAAIDev. Generate per-frame 3D "bullet-time" scenes from regular video in just 150ms on a single…
0
10
0
RT @ashmrz10: 🚀 Tired of waiting for your Gaussian-based scenes to fit dynamic inputs? ⏳ Wait no more! Check out our new paper and discover…
0
4
0
RT @huangjh_hjh: 📢Please check out our newest work on feed-forward reconstruction of dynamic monocular videos! With our bullet-time formula…
0
9
0
RT @ZGojcic: Reconstruct and explore monocular dynamic videos in real time! Interaction with your favorite video content is now possible w…
0
2
0
This is joint work with many great co-authors and could not be finished with all their diligence and devotion to the project! 🙌 @hx_liang95 @ashmrz10 @abtorralba @liuziwei7 @igilitschenski @FidlerSanja @CengizOztireli @HuanLing6 @ZGojcic @huangjh_hjh
0
0
7
BTimer retains a strong capability to reconstruct static scenes thanks to the bullet time formulation. The same kitchen-sink model trained on dynamic data, outperforms many baselines, including GS-LRM in Tanks & Temples datasets.
0
0
2
BTimer achieves SoTA performance on dynamic scenes, significantly reducing reconstruction time compared to existing approaches. We can reconstruct and render dynamic scenes in real time, surpassing many per-scene optimization-based methods in terms of both speed and quality.
0
0
2
We use curriculum training on a mix of static and dynamic data, enabling robust generalization across various scenarios. Training starts with low-resolution static scenes and progressively moves to higher resolution and dynamic data, increasing the complexity of the task.
0
0
1
We also design a Novel Time Enhancer (NTE) module that synthesizes intermediate frames for smoother reconstruction. NTE is crucial for maintaining continuity in fast-moving sequences and enables us to reconstruct the scene at timestamps for which we don't have an input frame.
0
0
4
BTimer is built on a 24-layer ViT backbone. We tokenize input frames and add two types of time embeddings: one for the input frame's timestamp, and the other for bullet timestamp. The model outputs pixel-aligned Gaussians that reconstruct the 3D scene at the bullet timestamp.
0
0
0