Qiang.W

Computer Vision, 3D Vision, Deep Learning

[Qiang.W @supertoday]

@GroundedSI Seem your SLAM diverged during this sample

[Qiang.W @supertoday]

@Ace3Zi @toolshed_labs @notgiannei Get it. It is still a good starting point and good luck

[Qiang.W @supertoday]

@Ace3Zi @toolshed_labs @notgiannei I don't get it. The results on LaMaria and your blogpost differ a lot. what happened?

[Qiang.W @supertoday]

@CSProfKGD The old days when we used to use pytorch

[Qiang.W @supertoday]

@CSProfKGD well
https://t.co/xIOuIql4Gl

[Kirill Mazur @makezur]

AgentSTAR is an agentic method for both reconstruction and tracking of articulated objects from monocular videos.

We can now track through rapid motion, severe occlusion and thin objects! 

There’s no spoon but we can still track it 🥄

[Qiang.W @supertoday]

@Ace3Zi @toolshed_labs @notgiannei Can wait for the paper. Is there an ETA for the paper?

[Qiang.W @supertoday]

@pesarlin @eccvconf @mapo1 Well deserved

[Phillip Isola @phillip_isola]

Recently, there have been a lot of impressive demos of AI agents, like Claude, controlling robots.

I wrote a short blog post with my thoughts on the advent of these "robot-use agents."

https://t.co/5bUjZlZxCi

I think it's an important change in the trajectory of robotics!

[Qiang.W @supertoday]

The transition model is trained on MSLS sequences and evaluated on similarly constrained ground trajectories, so generalization to handheld, pedestrian, or more general 6-DoF motion remains an interesting open question.

[Qiang.W @supertoday]

TRAIL #ECCV2026 sequential VPR: instead of assuming an ordered reference sequence, it works with an unordered geo-tagged database. Per-frame MegaLoc retrieval is combined with a learned transition score from DINO feature correlations, propagated over time with CRF-style filter.

[Qiang.W @supertoday]

paper:
https://t.co/99HZ2Ura6R

[Qiang.W @supertoday]

project:
https://t.co/jH2AjPjyCQ
paper:
https://t.co/Ib2ftLcqlJ

[Qiang.W @supertoday]

RoGe: Novel View Synthesis via End-to-End Implicit Reconstruction and Generation from Xiaomi EV
One model unifies reconstruction and generation. Given a few posed images and a camera trajectory, it synthesizes a coherent video along the traj.
similar to atlas from worldlabs

[Qiang.W @supertoday]

Pi3-like formulation, relative pose instead of global pose, rotation refiner.

[Qiang.W @supertoday]

Interesting work from Amap (高德)

[Zhenjun Zhao @zhenjun_zhao]

Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction

Jiarong Han, Jincheng Xiong, Yuzhou Liu, Linzhe Shi, Changjie Wu, Ning Guo, Mu Xu, Hang Zhang, Ming Qian

tl;dr: fixed 12-frame local context

https://t.co/6RelnH8vPb

[Qiang.W @supertoday]

It works on fixed 12 frames without no recurrent state or external memory but the drift is low.

[Qiang.W @supertoday]

The tricky part is 3D Foudantion models can work with only videos. But to add inertial cues such as this paper, one have to do camera–IMU extrinsics calibration .etc which narrows the usage. Besides classic visual inertial odometry or VINS methods works well when calibrated.

[Zhenjun Zhao @zhenjun_zhao]

VI3: Grounding Pretrained 3D Foundation Models with Inertial Cues

Ernesto Lozano, Alberto Jaenal, @jcivera

tl;dr: IMU + 3D foundation model (3DFM)->metric 3DFM

https://t.co/oB2jXrSuuU

[Qiang.W @supertoday]

It succeeds reconstructs my "challenging" sequence as well.

[Qiang.W @supertoday]

Interesting work from Amap (高德)

[Keunhong Park @KeunhongP]

When we say Atlas has pixel-perfect camera control, we mean it.

Atlas precisely follows input camera parameters, including non-planar projections such as the Brown-Conrady distortion model and the Kannala-Brandt fish-eye model.

@BhamidipatiPan1 invented a novel method of camera conditioning and it works beautifully.

🧵 [1/N]