[FFmpeg-devel] [PATCH] avfilter: add findandcover filter
Michael Niedermayer
michaelni at gmx.at
Thu Apr 30 14:08:00 CEST 2015
On Thu, Apr 30, 2015 at 09:49:27AM +0200, Clément Bœsch wrote:
> On Wed, Apr 29, 2015 at 06:00:54PM +0200, Michael Niedermayer wrote:
> > Signed-off-by: Michael Niedermayer <michaelni at gmx.at>
[...]
> > + c = (n*oh_sum_v - o_sum_v*(int64_t)h_sum_v) / (sqrt(o_sigma)*sqrt(h_sigma));
>
> not using sqrt(o_sigma * h_sigma) for precision or overflow concerns?
overflow
[...]
> > + for (y=0; y<h; y++) {
> > + for (x=0; x<w; x++) {
> > + data[x] = src[x];
> > + }
>
> > + data += in->linesize[p];
> > + src += foc->cover_frame->linesize[p];
>
> i've seen gcc derping a lot on this; you might want to check if taking
> these dereferencing out of the loop helps
i dont think that at this point of this loop it can have any overall
relevant effect
>
> > + }
> > + }
> > +}
> > +static void blur(FOCContext *foc, AVFrame *in, int offx, int offy)
> > +{
> > + int x, y, p;
> > +
> > + for (p=0; p<3; p++) {
>
> > + int ox = offx>>!!p;
> > + int oy = offy>>!!p;
>
> FF_CEIL_RSHIFT?
no that would be wrong here
others changed to use FF_CEIL_RSHIF
>
> > + int stride = in->linesize[p];
> > + uint8_t *data = in->data[p] + ox + oy * stride;
> > + int w = foc->obj_frame->width >> !!p;
> > + int h = foc->obj_frame->height >> !!p;
> > + int iw = in->width >> !!p;
> > + int ih = in->height >> !!p;
> > + for (y=0; y<h; y++) {
> > + for (x=0; x<w; x++) {
> > + int c = 0;
> > + int s = 0;
> > + if (ox) {
> > + int scale = 65536 / (x + 1);
> > + s += data[-1 + y*stride] * scale;
> > + c += scale;
> > + }
> > + if (oy) {
> > + int scale = 65536 / (y + 1);
> > + s += data[x - stride] * scale;
> > + c += scale;
> > + }
> > + if (ox + w < iw) {
> > + int scale = 65536 / (w - x);
> > + s += data[w + y*stride] * scale;
> > + c += scale;
> > + }
> > + if (oy + h < ih) {
> > + int scale = 65536 / (h - y);
> > + s += data[x + h*stride] * scale;
> > + c += scale;
> > + }
> > + data[x + y*stride] = (s + (c>>1)) / c;
> > + }
> > + }
> > + }
> > +}
> > +
> > +static int filter_frame(AVFilterLink *inlink, AVFrame *in)
> > +{
> > + AVFilterContext *ctx = inlink->dst;
> > + FOCContext *foc = ctx->priv;
> > + float best_score;
> > + int best_x, best_y;
> > + int i;
> > +
> > + foc->haystack_frame[0] = av_frame_clone(in);
> > + for (i=1; i<foc->mipmaps; i++) {
> > + foc->haystack_frame[i] = downscale(foc->haystack_frame[i-1]);
> > + }
> > +
> > + best_score = search(foc, 0, 0,
> > + FFMAX(foc->xmin, foc->last_x - 8),
> > + FFMIN(foc->xmax, foc->last_x + 8),
> > + FFMAX(foc->ymin, foc->last_y - 8),
> > + FFMIN(foc->ymax, foc->last_y + 8),
> > + &best_x, &best_y, 1.0);
> > +
> > + best_score = search(foc, 0, foc->mipmaps - 1, foc->xmin, foc->xmax, foc->ymin, foc->ymax,
> > + &best_x, &best_y, best_score);
> > +
> > + for (i=0; i<MAX_MIPMAPS; i++) {
> > + av_frame_free(&foc->haystack_frame[i]);
> > + }
> > +
> > + if (best_score > foc->threshold) {
> > + return ff_filter_frame(ctx->outputs[0], in);
> > + }
> > +
>
> > + av_log(ctx, AV_LOG_DEBUG, "Found at %d %d score %f\n", best_x, best_y, best_score);
>
> found what?
the filter is called "find and cover" so i think "Found" is clear
>
> > + foc->last_x = best_x;
> > + foc->last_y = best_y;
> > +
> > + av_frame_make_writable(in);
> > +
> > + if (foc->mode == MODE_BLUR) {
> > + blur (foc, in, best_x, best_y);
> > + } else {
> > + cover(foc, in, best_x, best_y);
> > + }
> > + return ff_filter_frame(ctx->outputs[0], in);
> > +}
> > +
> > +static av_cold void uninit(AVFilterContext *ctx)
> > +{
> > + FOCContext *foc = ctx->priv;
> > + int i;
> > +
> > + for (i=0; i<MAX_MIPMAPS; i++) {
>
> nit: pleasefixthestyle
google says
mipmap has 287,000 result
mip map has 57,500 results
>
> > + av_frame_free(&foc->needle_frame[i]);
> > + av_frame_free(&foc->haystack_frame[i]);
> > + }
> > +
> > + if (foc->obj_frame)
> > + av_freep(&foc->obj_frame->data[0]);
> > + if (foc->cover_frame)
> > + av_freep(&foc->cover_frame->data[0]);
> > + av_frame_free(&foc->obj_frame);
> > +}
> > +
> > +static av_cold int init(AVFilterContext *ctx)
> > +{
> > + FOCContext *foc = ctx->priv;
> > + int ret, i;
> > +
> > + if (!foc->obj_filename || (!foc->cover_filename && foc->mode == MODE_COVER)) {
> > + av_log(ctx, AV_LOG_ERROR, "object or cover filename not set\n");
> > + return AVERROR(EINVAL);
> > + }
> > +
> > + foc->obj_frame = av_frame_alloc();
> > + if (!foc->obj_frame)
> > + return AVERROR(ENOMEM);
> > +
>
> > + if ((ret = ff_load_image(foc->obj_frame->data, foc->obj_frame->linesize,
> > + &foc->obj_frame->width, &foc->obj_frame->height,
> > + &foc->obj_frame->format, foc->obj_filename, ctx)) < 0)
> > + return ret;
>
> Can you use the dualinput mechanism to load the image instead? See
> paletteuse for a typical 1 frame load case. This will avoid a libavformat
> and libavcodec dependency to the filter (which you forgot to add to
> configure), as well as clumsiness with the file path into the filtergraph.
iam a bit too lazy to implement that, also there are 3 inputs (the
input video, the object to find and the replacement object to paint
over, the last is optional)
ill post a patch with all other issues fixed in a moment
if you want you can change this one remaining and push it
thanks
[...]
--
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
Republics decline into democracies and democracies degenerate into
despotisms. -- Aristotle
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20150430/555d7daf/attachment.asc>
More information about the ffmpeg-devel
mailing list