Making temporal filtering bidirectional?

[Steen4]

I'm not really a technically-minded person, but I've been wondering lately: would it be feasible to create a bidirectional temporal nr filter for virtualdub? Maybe buffer n number of frames around a target frame, process forward & backwards to the target frame, perform threshold comparison (to unprocessed frame?) between pixels in forward version of processed frame and in backwards version of processed frame (this should catch any egregious motion artifacts in the middle), and wherever in the frame the threshold is exceeded, use either information from unprocessed version of target frame or data interpolated from surrounding pixels, otherwise use blending of forward version and backwards version of processed frame. Although temporal filters usually run quickly compared to spatial filters, I would imagine this to be fairly slow; however, this type of processing, especially after an instance of something like Smart Smoother HiQ, would be quite impressive...and largely artifact-free...or am I missing something obvious? Don't hesitate to point out any flaws in my logic or approach. I know I'm askin' for trouble...

[ronnylov]

I thought that was exactly what temporal filters do, look at frames before and after current frame. I'm not sure with virtualdub but at least with avisynth you can do that. For instance the temporalsoften filter in avisynth can set number of frames to use and some thresholds.

[Steen4]

No, I think that the current temporal filters all buffer in one direction: n frames previous to the current frame up to the current frame. To see the action of a temporal filter (and its artifacts) more clearly, try this test: open up virtualdub, then load a video (preferably high-motion). Next, load any temporal filter (temp cleaner, temp smoother, dnr, motion blur, etc) and step through the motion scenes, both backwards and forward. If the artifacts present in frame n are the same stepping forward as the artifacts present when stepping backwards, then the filter is processing bidirectionally. The filters I have seen in vdub for temporal processing only operate in one direction with no lookahead, however, so I was wondering whether this was because of the limitations of either vdub itself, the vdub filter sdk, or other filter-writing tools. If not, then why hasn't it been done? Too costly in terms of processing power/too slow? Memory hog? Results not worth the extra performance hit?