Body interpolation is the method of synthesizing in-between photos from a given set of photos. The approach is commonly used for temporal up-sampling to extend the refresh price of movies or to create sluggish movement results. These days, with digital cameras and smartphones, we frequently take a number of images inside a number of seconds to seize one of the best image. Interpolating between these “near-duplicate” images can result in partaking movies that reveal scene movement, usually delivering an much more pleasing sense of the second than the unique images.
Body interpolation between consecutive video frames, which regularly have small movement, has been studied extensively. Not like movies, nonetheless, the temporal spacing between near-duplicate images will be a number of seconds, with commensurately giant in-between movement, which is a serious failing level of current body interpolation strategies. Latest strategies try to deal with giant movement by coaching on datasets with excessive movement, albeit with restricted effectiveness on smaller motions.
In “FILM: Body Interpolation for Giant Movement”, printed at ECCV 2022, we current a way to create top quality slow-motion movies from near-duplicate images. FILM is a brand new neural community structure that achieves state-of-the-art ends in giant movement, whereas additionally dealing with smaller motions properly.
![]() |
FILM interpolating between two near-duplicate images to create a sluggish movement video. |
FILM Mannequin Overview
The FILM mannequin takes two photos as enter and outputs a center picture. At inference time, we recursively invoke the mannequin to output in-between photos. FILM has three elements: (1) A characteristic extractor that summarizes every enter picture with deep multi-scale (pyramid) options; (2) a bi-directional movement estimator that computes pixel-wise movement (i.e., flows) at every pyramid degree; and (3) a fusion module that outputs the ultimate interpolated picture. We practice FILM on common video body triplets, with the center body serving because the ground-truth for supervision.
![]() |
An ordinary characteristic pyramid extraction on two enter photos. Options are processed at every degree by a sequence of convolutions, that are then downsampled to half the spatial decision and handed as enter to the deeper degree. |
Scale-Agnostic Characteristic Extraction
Giant movement is usually dealt with with hierarchical movement estimation utilizing multi-resolution characteristic pyramids (proven above). Nevertheless, this methodology struggles with small and fast-moving objects as a result of they will disappear on the deepest pyramid ranges. As well as, there are far fewer accessible pixels to derive supervision on the deepest degree.
To beat these limitations, we undertake a characteristic extractor that shares weights throughout scales to create a “scale-agnostic” characteristic pyramid. This characteristic extractor (1) permits the usage of a shared movement estimator throughout pyramid ranges (subsequent part) by equating giant movement at shallow ranges with small movement at deeper ranges, and (2) creates a compact community with fewer weights.
Particularly, given two enter photos, we first create a picture pyramid by successively downsampling every picture. Subsequent, we use a shared U-Web convolutional encoder to extract a smaller characteristic pyramid from every picture pyramid degree (columns within the determine beneath). Because the third and last step, we assemble a scale-agnostic characteristic pyramid by horizontally concatenating options from totally different convolution layers which have the identical spatial dimensions. Be aware that from the third degree onwards, the characteristic stack is constructed with the identical set of shared convolution weights (proven in the identical coloration). This ensures that every one options are related, which permits us to proceed to share weights within the subsequent movement estimator. The determine beneath depicts this course of utilizing 4 pyramid ranges, however in observe, we use seven.
Bi-directional Circulation Estimation
After characteristic extraction, FILM performs pyramid-based residual move estimation to compute the flows from the yet-to-be-predicted center picture to the 2 inputs. The move estimation is finished as soon as for every enter, ranging from the deepest degree, utilizing a stack of convolutions. We estimate the move at a given degree by including a residual correction to the upsampled estimate from the following deeper degree. This method takes the next as its enter: (1) the options from the primary enter at that degree, and (2) the options of the second enter after it’s warped with the upsampled estimate. The identical convolution weights are shared throughout all ranges, apart from the 2 most interesting ranges.
Shared weights permit the interpretation of small motions at deeper ranges to be the identical as giant motions at shallow ranges, boosting the variety of pixels accessible for big movement supervision. Moreover, shared weights not solely allow the coaching of highly effective fashions that will attain a better peak signal-to-noise ratio (PSNR), however are additionally wanted to allow fashions to suit into GPU reminiscence for sensible functions.
![]() |
![]() |
The affect of weight sharing on picture high quality. Left: no sharing, Proper: sharing. For this ablation we used a smaller model of our mannequin (referred to as FILM-med within the paper) as a result of the total mannequin with out weight sharing would diverge because the regularization advantage of weight sharing was misplaced. |
Fusion and Body Era
As soon as the bi-directional flows are estimated, we warp the 2 characteristic pyramids into alignment. We receive a concatenated characteristic pyramid by stacking, at every pyramid degree, the 2 aligned characteristic maps, the bi-directional flows and the enter photos. Lastly, a U-Web decoder synthesizes the interpolated output picture from the aligned and stacked characteristic pyramid.
Loss Capabilities
Throughout coaching, we supervise FILM by combining three losses. First, we use the absolute L1 distinction between the anticipated and ground-truth frames to seize the movement between enter photos. Nevertheless, this produces blurry photos when used alone. Second, we use perceptual loss to enhance picture constancy. This minimizes the L1 distinction between the ImageNet pre-trained VGG-19 options extracted from the anticipated and floor reality frames. Third, we use Model loss to attenuate the L2 distinction between the Gram matrix of the ImageNet pre-trained VGG-19 options. The Model loss permits the community to provide sharp photos and reasonable inpaintings of huge pre-occluded areas. Lastly, the losses are mixed with weights empirically chosen such that every loss contributes equally to the whole loss.
Proven beneath, the mixed loss enormously improves sharpness and picture constancy when in comparison with coaching FILM with L1 loss and VGG losses. The mixed loss maintains the sharpness of the tree leaves.
![]() |
FILM’s mixed loss features. L1 loss (left), L1 plus VGG loss (center), and Model loss (proper), exhibiting important sharpness enhancements (inexperienced field). |
Picture and Video Outcomes
We consider FILM on an inside near-duplicate images dataset that displays giant scene movement. Moreover, we examine FILM to latest body interpolation strategies: SoftSplat and ABME. FILM performs favorably when interpolating throughout giant movement. Even within the presence of movement as giant as 100 pixels, FILM generates sharp photos in step with the inputs.
![]() |
Body interpolation with SoftSplat (left), ABME (center) and FILM (proper) exhibiting favorable picture high quality and temporal consistency. |
![]() |
Conclusion
We introduce FILM, a big movement body interpolation neural community. At its core, FILM adopts a scale-agnostic characteristic pyramid that shares weights throughout scales, which permits us to construct a “scale-agnostic” bi-directional movement estimator that learns from frames with regular movement and generalizes properly to frames with giant movement. To deal with vast disocclusions attributable to giant scene movement, we supervise FILM by matching the Gram matrix of ImageNet pre-trained VGG-19 options, which ends up in reasonable inpainting and crisp photos. FILM performs favorably on giant movement, whereas additionally dealing with small and medium motions properly, and generates temporally clean top quality movies.
Attempt It Out Your self
You possibly can check out FILM in your images utilizing the supply code, which is now publicly accessible.
Acknowledgements
We wish to thank Eric Tabellion, Deqing Solar, Caroline Pantofaru, Brian Curless for his or her contributions. We thank Marc Comino Trinidad for his contributions on the scale-agnostic characteristic extractor, Orly Liba and Charles Herrmann for suggestions on the textual content, Jamie Aspinall for the imagery within the paper, Dominik Kaeser, Yael Pritch, Michael Nechyba, William T. Freeman, David Salesin, Catherine Wah, and Ira Kemelmacher-Shlizerman for help. Due to Tom Small for creating the animated diagram on this put up.