FlowCam: Training Generalizable 3D Radiance Fields without Camera Poses via Pixel-Aligned Scene Flow

Reconstruction of 3D neural fields from posed images has emerged as a promising method for self-supervised representation learning. The key challenge preventing the deployment of these 3D scene learners on large-scale video data is their dependence on precise camera poses from structure-from-motion, which is prohibitively expensive to run at scale. We propose a method that jointly reconstructs camera poses and 3D neural scene representations online and in a single forward pass. We estimate poses by first lifting frame-to-frame optical flow to 3D scene flow via differentiable rendering, preserving locality and shift-equivariance of the image processing backbone. SE(3) camera pose estimation is then performed via a weighted least-squares fit to the scene flow field. This formulation enables us to jointly supervise pose estimation and a generalizable neural scene representation via re-rendering the input video, and thus, train end-to-end and fully self-supervised on real-world video datasets. We demonstrate that our method performs robustly on diverse, real-world video, notably on sequences traditionally challenging to optimization-based pose estimation techniques.

Paper

Similar papers

Peer review

Reviewer 52at6/10 · confidence 5/52023-06-29

Summary

This paper presents a generalizable framework for estimating both the NeRF model of the scene and the camera pose for video sequence simultiunusly. The proposed method first uses the pretrained model to predict the optical flow for each video frame. Then, the optical flow is lifted up to the monodepth to scene flow. The camera poses are solved by SVD with scene flow. After computing the camera pose, the nerf model is further fine-tuned for a better result.

Strengths

- a light step of unposed nerf in RGB video sequence setting. - the quality of the rendered RGB image looks good after fine-tune. - well written, easy to read and follow.

Weaknesses

- Statement - Line 27-31 - Although the manuscript does acknowledge certain NeRF-based SLAM methods for their ability to provide dense 3D reconstruction, relevant references [1,2,3] are conspicuously absent. Additionally, some NeRF-based SLAM methods that utilize prior for pose estimation have been glossed over. - Line 34 - The cited works do not adequately represent the body of literature since both methods focus exclusively on depth map estimation. This oversight should be rectified by including relevant citations [4,5]. - Line 143 - wrong format - Motivation - The motivation for introducing single view generalizable nerf for relative pose estimation is unclear and wired. - Line 142 - the definition of the x_t. In equ2, x_i is the sampled points along the ray. Does x_t is the S(r) in equ2? - while pixelnerf can use only a single view to render a novel view, it is much like a mono-depth estimation module. Without optical flow, an alternating solution uses iNeRF to align two single views point cloud/nerf, minimizing photometric loss. However, even most SOTA mono-depth estimation methods are insufficient to estimate accurate relative pose due to cross-view scale misalignment. Some unposed/posed Nerf only use mono-depth as a reg term. This is why the proposed method **HAS TO** introduce the out-of-shelf optical flow estimation for pose estimation. If the single view pixelnerf is removed from the proposed method, optical flow, as the correspondences, is enough to estimate the essential matrix/relative pose between two neighbor frames. Furthermore, the deep-sfm module, like banet, droidslam, deepv3d, all of these methods take multiview images to estimate both multiview depth map and the camera pose at the same time. The author should carefully address this issue during the rebuttal. - the motivation of flow confidence weights. - Line 143 - This part could be easily solved by bi-direction consistency check with optical flow. Introducing another pretrained network for weight prediction will make the generalization issue worse. - Figure 9 does not show the results with dynamic objects. - Line 178 - RGB VO/SLAM confronts scale ambiguity problem. The proposed method doesn't handle it at all. - Line 194 - where is the definition of N, J? - Line 203 Equ8 - The mono-depth is enforced by this loss term during fine-tune. As mentioned above, the optical flow for relative pose estimation is enough. - This loss term does not handle singular solutions. - Experiment - Line 211 - the training detail is missing. - the proposed method cannot run on a long video clip. - Most of the experiments show the object-center case. It will be great to see the results on the inside-out case either. - Although the authors mention that they do not claim sota accuracy of pose estimation, the metrics of the pose are still necessary to be reported on the main paper, including droid slam, droid vo, orb slam, etc. - Line 224 - test-time optimization on the weight of CNN and MLP on a single video clip in a self-supervised manner is extremely hard. The problem will become worse if the precomputed optical flow is removed. The author should provide more detail about this section. - Line 250 - more baseline methods are required, such as nicer-slam[2], dim-slam[3], droid-slam. - some unposed nerf methods with mono depth estimation module, such as nope-nerf[6] - and another test-time optization method, such as CVD[4], RCVD[5]. - The pose trajectories in the figure are misaligned in both the main paper and the supplementary, which are very confusing. - reference - [1] Sucar, E., Liu, S., Ortiz, J., & Davison, A. J. (2021). iMAP: Implicit mapping and positioning in real-time. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 6229-6238). - [2] Zhu, Z., Peng, S., Larsson, V., Cui, Z., Oswald, M. R., Geiger, A., & Pollefeys, M. (2023). Nicer-slam: Neural implicit scene encoding for rgb slam. arXiv preprint arXiv:2302.03594. - [3] Li, H., Gu, X., Yuan, W., Yang, L., Dong, Z., & Tan, P. (2023). Dense RGB SLAM with Neural Implicit Maps. The Eleventh International Conference on Learning Representations(ICLR) - [4] Luo, X., Huang, J. B., Szeliski, R., Matzen, K., & Kopf, J. (2020). Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4), 71-1. - [5] Kopf, J., Rong, X., & Huang, J. B. (2021). Robust consistent video depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 1611-1621). - [6] Bian, W., Wang, Z., Li, K., Bian, J. W., & Prisacariu, V. A. (2023). Nope-nerf: Optimising neural radiance field with no pose prior. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 4160-4169).

Questions

I am happy to adjust the rating if the author properly addresses the concerns mentioned during the rebuttal. ---- after rebuttal: I adjust the rate to the positive side.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Yes.

Reviewer kV1r7/10 · confidence 4/52023-06-30

Summary

This paper proposes a method to address the challenge of reconstructing 3D neural fields from images and learns them in a self-supervised manner. The main contribution of this method is the joint reconstruction of camera poses and 3D neural scene representations within a single forward pass.

Strengths

1. This paper achieves state-of-the-art performance in its specific setting. 2. The writing in this paper is clear. 3. This paper addresses an important problem. It overcomes the previous limitation in training generalized Neural Radiance Fields that require posed videos. It enables direct training on unposed video data. 4. This paper is innovative in its methodology and no one has utilized similar approaches to address this problem before.

Weaknesses

1. The paper lacks clarity in explaining the ablation study, such as the incomplete caption for Table 3. Moreover, the abbreviations used in the table are not defined, making it unclear what they refer to. 2. The paper lacks ablation experiments on not learning intrinsic camera parameters.

Questions

1. I cannot understand how this paper addresses the generalization issue. Why does PixelNeRF have good generalization capabilities in monocular depth prediction? 2. How does this paper address the problem of scale ambiguity in monocular depth estimation?

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

This paper discusses its limitations. However, I would have expected to see more failure cases from internet data.

Reviewer 3ZaF5/10 · confidence 5/52023-07-06

Summary

This paper proposes a general method for 3D neural scene reconstruction and camera pose estimation from a video sequence. The method takes a set of video frames as input and outputs the re-rendered video frames and estimated camera poses. The method is based on PixelNeRF, which is a general NeRF method that takes extracted image features as input. The method first uses single-view PixelNeRF to compute per-frame point clouds. Then, the frame-to-frame 3D scene flow is estimated by leveraging the estimated 2D optical flow. The 3D scene flow is then used to constrain the frame-to-frame camera poses, which are solved by minimizing a weighted least-squares problem. Two losses are used to supervise the training: the RGB loss, which forces the PixelNeRF-based rendered images to be close to the input images; and the flow loss, which makes the projection offset of neighbor frames' point clouds close to the 2D optical flow.

Strengths

1. A nice solution to learn a general NeRF representation for multiview images without camer pose information. 2. The quality of the rendering results for testing multi-view images are impressive.

Weaknesses

The experiments can be improved to further verified the advantage of the proposed method: 1. Only one quantitative pose estimation comparison is shown, there is no quantitative comparisons with BARF. Besides, Nope-NeRF is a similar method that takes sequent frames as input for pose estimation, it better have one more comparison with Nope-NeRF. 2. The quantitative results in Tab 1 are fine-tuned or not? It should be better to show both results.

Questions

Overall, the proposed method is reasonable, but I do have questions: 1. There should be a scale problem when estimating the per-view point cloud by single-view pixelNeRF. It seems that the Eq (4) didn't consider it, why? 2. The confidence weight is output by a network. Therefore, in my understanding, this network can implicitly handle the occlusion and specular by downweighting corresponding flow. But it seems that the Masked RGB images shown in Fig 9 mask out too much content (even for textured diffuse regions). Is this because of the inaccurate 2D optical flow estimation?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The proposed method learns a conditioned generaI NeRF network for multiview images. It is still limited since the images should be kept to accompany the learned model.

Reviewer oZ256/10 · confidence 4/52023-07-06

Summary

The paper proposes using scene flow to optimize for camera poses to produce generalizable 3D radiance field. The key contribution is the joint optimization of the camera poses and 3D neural scene representations in an single forward pass. The method has been evaluated on multiple datasets and performs well on datasets where traditional pose estimation fails.

Strengths

The motivation for using an joint optimization of camera poses and reconstruction has wide applications and easy to scale. As proposed by the current method multiple views of a scene inherently produce good correspondences, which leads to reasonable optical flow estimates. Using and incorporating optical flow into the 3D pose estimation and reconstruction helps overall reconstruction pipeline. Using optical flow for reconstruction and Neural radiance fields without cameras poses have been well independently. Combining them into a single formulation is still less explored. the current method lifts the optical flow to scene flow using neural scene representations and uses them to optimize. This is interesting direction to explore. The results shows that the method is able to perform better than traditional pose estimation methods like ORB-SLAM in some of their failure cases.

Weaknesses

The current method proposes a tradeoff by stating instead of using camera poses first and neural radiance field later as a two stage process, joint optimization is a more promising direction. but using joint optimization adds additional optimization constraints which have not been extensively discussed. It would be nice to see what are the failure cases of using the joint optimization compared to the two-stage optimization process. The camera pose estimation has only been evaluated with ORB-SLAM3 or VidAE both of which do not incorporate neural radiance fields for 3D pose estimation. Recently there have been better methods which optimize for both radiance fields and pose simultaneously. having such baseline will give a better understanding of the accuracy of the algorithm. The number of frames used for error computation is very small. i.e. 20 frames and 200 frames. The advantage of the method states that the joint optimization is beneficial to scale easily. Showing the method accuracy compared to orb on larger datasets or sequences like

Questions

in Fig3 results were shown using BARF for pose estimation, why are there no quantitative results using the same?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Limitations have been well discussed.

Reviewer kV1r2023-08-18

Thanks for the rebuttal, which has well addressed my concerns. So I keep the positive side.

Reviewer 52at2023-08-19

Thanks for the rebuttal. Although I still have concerns about the `scale drifting`, which is not the `scale ambiguity issue responded to by the author, most concerns are properly addressed during the rebuttal. Some missing Nerf-slam-based methods still need a proper reference.

Authorsrebuttal2023-08-21

Follow-up response on scale drift

Thanks for responding and again for the detailed review, as well as for the the clarification on the type of scale discrepancy referenced. Since our method makes frame-to-frame estimates, as opposed to performing multi-frame optimization, indeed scale drift error could potentially accumulate in the feedforward setting. Adding a scale parameter to the Procrustes could address this, but we found that forcing the model to use a single scale yielded more stable training (Tab. 2 of author response PDF). Enforcing a single scale also has the benefit of more temporally consistent depth estimates, as opposed to the scale-invariant depth estimated commonly by monocular depth networks which are notoriously difficult to normalize correctly. Multi-frame feedforward optimization is exciting future work. We will make sure to add the NeRF-SLAM methods mentioned.

Reviewer 3ZaF2023-08-21

Thanks for clarifying my questions. Please incorporate the materials in the rebuttal into the revision

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC