MFGS
A feed-forward multi-view feature encoding framework for 3D Gaussian Splatting that embeds multi-view feature encoding into the feed-forward pipeline, enabling end-to-end joint optimization of bitrate and distortion.
Timeline (last 90 days)
The MFGS framework embeds multi-view feature encoding directly into the feed-forward 3DGS pipeline, achieving up to 99.86% bitrate savings (BD-rate) on the three datasets RE10K, ACID and DL3DV, with about a 12x end-to-end speedup.
MFGS's core method is to compress the multi-view features extracted from the input images rather than the explicit Gaussian representation: the encoder extracts compact features, and the decoder directly predicts Gaussian attributes to complete rendering.
On the three datasets RE10K, ACID and DL3DV, MFGS saves 99.76% to 99.86% of bitrate as measured by BD-rate (the average percentage of bitrate saved relative to the baseline method at the same reconstruction quality).
MFGS reduces total GMACs by 52% to 55%, meaning the compute required by the model for the same task is nearly halved.
MFGS introduces an importance-guided Gaussian selection mechanism that allows the receiver to flexibly adjust the number of Gaussians used for rendering from the same bitstream, without regenerating or re-encoding the bitstream.
MFGS's bitrate savings are highly consistent across the three datasets, indicating that the gains come from architectural improvements rather than adaptation to a specific dataset.
In multi-view 3D reconstruction pipelines, compression should not happen after representation generation but should proceed in sync with representation learning, so that the model spontaneously learns an intermediate representation that both renders well and compresses well.
All Facts (7)
The MFGS framework embeds multi-view feature encoding directly into the feed-forward 3DGS pipeline, achieving up to 99.86% bitrate savings (BD-rate) on the three datasets RE10K, ACID and DL3DV, with about a 12x end-to-end speedup.
50%UnverifiedMFGS's core method is to compress the multi-view features extracted from the input images rather than the explicit Gaussian representation: the encoder extracts compact features, and the decoder directly predicts Gaussian attributes to complete rendering.
50%UnverifiedOn the three datasets RE10K, ACID and DL3DV, MFGS saves 99.76% to 99.86% of bitrate as measured by BD-rate (the average percentage of bitrate saved relative to the baseline method at the same reconstruction quality).
50%UnverifiedMFGS reduces total GMACs by 52% to 55%, meaning the compute required by the model for the same task is nearly halved.
50%UnverifiedMFGS introduces an importance-guided Gaussian selection mechanism that allows the receiver to flexibly adjust the number of Gaussians used for rendering from the same bitstream, without regenerating or re-encoding the bitstream.
50%UnverifiedMFGS's bitrate savings are highly consistent across the three datasets, indicating that the gains come from architectural improvements rather than adaptation to a specific dataset.
50%UnverifiedIn multi-view 3D reconstruction pipelines, compression should not happen after representation generation but should proceed in sync with representation learning, so that the model spontaneously learns an intermediate representation that both renders well and compresses well.
50%