[PATCH] avfilter/vf_bwdif: Remove undesireable spatial preference logic
bwdif inherited this check from yadif, which was originally supposed to prefer the spatial predictor if the temporal predictor was too far off. However, the core bwdif algorithm already accounts for the spatial predictor, so this additional check actually ends up preferring a worse value, reducing the overall quality. This was found by cyanreg while writing bwdif_vulkan, and the visual improvement is pretty dramatic in some samples. If we agree that this change is desirable, we should update all implementations. Signed-off-by: Philip Langdale <philipl@overt.org> --- libavfilter/vf_bwdif.c | 5 ----- 1 file changed, 5 deletions(-) diff --git a/libavfilter/vf_bwdif.c b/libavfilter/vf_bwdif.c index 65c617ebb3..441bb11e7b 100644 --- a/libavfilter/vf_bwdif.c +++ b/libavfilter/vf_bwdif.c @@ -106,11 +106,6 @@ typedef struct ThreadData { interpol = (c + e) >> 1; #define FILTER2() \ - if (interpol > d + diff) \ - interpol = d + diff; \ - else if (interpol < d - diff) \ - interpol = d - diff; \ - \ dst[0] = av_clip(interpol, 0, clip_max); \ } \ \ -- 2.37.2
Hi Philip, Philip Langdale <philipl@overt.org> schrieb am Fr., 24. März 2023, 23:21:
bwdif inherited this check from yadif, which was originally supposed to prefer the spatial predictor if the temporal predictor was too far off.
However, the core bwdif algorithm already accounts for the spatial predictor, so this additional check actually ends up preferring a worse value, reducing the overall quality.
This was found by cyanreg while writing bwdif_vulkan, and the visual improvement is pretty dramatic in some samples. If we agree that this change is desirable, we should update all implementations.
Signed-off-by: Philip Langdale <philipl@overt.org> --- libavfilter/vf_bwdif.c | 5 ----- 1 file changed, 5 deletions(-)
diff --git a/libavfilter/vf_bwdif.c b/libavfilter/vf_bwdif.c index 65c617ebb3..441bb11e7b 100644 --- a/libavfilter/vf_bwdif.c +++ b/libavfilter/vf_bwdif.c @@ -106,11 +106,6 @@ typedef struct ThreadData { interpol = (c + e) >> 1;
#define FILTER2() \ - if (interpol > d + diff) \ - interpol = d + diff; \ - else if (interpol < d - diff) \ - interpol = d - diff; \ - \
Removing this will make lower thirds and other graphic jump up and down each frame. It is the main improvement over w3fdif that I have ported from yadif. Can you provide samples including still graphics that are improved with this patch? Regards, Thomas
On Sat, 25 Mar 2023 00:02:03 +0100 Thomas Mundt <tmundt75@gmail.com> wrote:
Hi Philip,
Philip Langdale <philipl@overt.org> schrieb am Fr., 24. März 2023, 23:21:
bwdif inherited this check from yadif, which was originally supposed to prefer the spatial predictor if the temporal predictor was too far off.
However, the core bwdif algorithm already accounts for the spatial predictor, so this additional check actually ends up preferring a worse value, reducing the overall quality.
This was found by cyanreg while writing bwdif_vulkan, and the visual improvement is pretty dramatic in some samples. If we agree that this change is desirable, we should update all implementations.
Signed-off-by: Philip Langdale <philipl@overt.org> --- libavfilter/vf_bwdif.c | 5 ----- 1 file changed, 5 deletions(-)
diff --git a/libavfilter/vf_bwdif.c b/libavfilter/vf_bwdif.c index 65c617ebb3..441bb11e7b 100644 --- a/libavfilter/vf_bwdif.c +++ b/libavfilter/vf_bwdif.c @@ -106,11 +106,6 @@ typedef struct ThreadData { interpol = (c + e) >> 1;
#define FILTER2() \ - if (interpol > d + diff) \ - interpol = d + diff; \ - else if (interpol < d - diff) \ - interpol = d - diff; \ - \
Removing this will make lower thirds and other graphic jump up and down each frame. It is the main improvement over w3fdif that I have ported from yadif. Can you provide samples including still graphics that are improved with this patch?
Hi Thomas, Sorry for this thread going unanswered for so long. With the Vulkan hwaccel stuff sorted out, I wanted to come back to this. I discussed more with Lynne and we're no longer convinced this change is obviously desirable. You are right about the jumping behaviour - and although there are samples (eg: the burosch ones on samples.ffmpeg.org) which look better with the change, I don't think that's something to over index on. Thanks, --phil
Jun 11, 2023, 04:53 by philipl@overt.org:
On Sat, 25 Mar 2023 00:02:03 +0100 Thomas Mundt <tmundt75@gmail.com> wrote:
Hi Philip,
Philip Langdale <philipl@overt.org> schrieb am Fr., 24. März 2023, 23:21:
bwdif inherited this check from yadif, which was originally supposed to prefer the spatial predictor if the temporal predictor was too far off.
However, the core bwdif algorithm already accounts for the spatial predictor, so this additional check actually ends up preferring a worse value, reducing the overall quality.
This was found by cyanreg while writing bwdif_vulkan, and the visual improvement is pretty dramatic in some samples. If we agree that this change is desirable, we should update all implementations.
Signed-off-by: Philip Langdale <philipl@overt.org> --- libavfilter/vf_bwdif.c | 5 ----- 1 file changed, 5 deletions(-)
diff --git a/libavfilter/vf_bwdif.c b/libavfilter/vf_bwdif.c index 65c617ebb3..441bb11e7b 100644 --- a/libavfilter/vf_bwdif.c +++ b/libavfilter/vf_bwdif.c @@ -106,11 +106,6 @@ typedef struct ThreadData { interpol = (c + e) >> 1;
#define FILTER2() \ - if (interpol > d + diff) \ - interpol = d + diff; \ - else if (interpol < d - diff) \ - interpol = d - diff; \ - \
Removing this will make lower thirds and other graphic jump up and down each frame. It is the main improvement over w3fdif that I have ported from yadif. Can you provide samples including still graphics that are improved with this patch?
Hi Thomas,
Sorry for this thread going unanswered for so long. With the Vulkan hwaccel stuff sorted out, I wanted to come back to this. I discussed more with Lynne and we're no longer convinced this change is obviously desirable. You are right about the jumping behaviour - and although there are samples (eg: the burosch ones on samples.ffmpeg.org) which look better with the change, I don't think that's something to over index on.
I don't recall agreeing that this is desirable. bwdif_vulkan still looks better than bwdif or bwdif_cuda (and it's got proper edge handling). I just forgot to investigate this, that's all.
Lynne <dev@lynne.ee> schrieb am So., 11. Juni 2023, 20:11:
Jun 11, 2023, 04:53 by philipl@overt.org:
On Sat, 25 Mar 2023 00:02:03 +0100 Thomas Mundt <tmundt75@gmail.com> wrote:
Hi Philip,
Philip Langdale <philipl@overt.org> schrieb am Fr., 24. März 2023, 23:21:
bwdif inherited this check from yadif, which was originally supposed to prefer the spatial predictor if the temporal predictor was too far off.
However, the core bwdif algorithm already accounts for the spatial predictor, so this additional check actually ends up preferring a worse value, reducing the overall quality.
This was found by cyanreg while writing bwdif_vulkan, and the visual improvement is pretty dramatic in some samples. If we agree that this change is desirable, we should update all implementations.
Signed-off-by: Philip Langdale <philipl@overt.org> --- libavfilter/vf_bwdif.c | 5 ----- 1 file changed, 5 deletions(-)
diff --git a/libavfilter/vf_bwdif.c b/libavfilter/vf_bwdif.c index 65c617ebb3..441bb11e7b 100644 --- a/libavfilter/vf_bwdif.c +++ b/libavfilter/vf_bwdif.c @@ -106,11 +106,6 @@ typedef struct ThreadData { interpol = (c + e) >> 1;
#define FILTER2() \ - if (interpol > d + diff) \ - interpol = d + diff; \ - else if (interpol < d - diff) \ - interpol = d - diff; \ - \
Removing this will make lower thirds and other graphic jump up and down each frame. It is the main improvement over w3fdif that I have ported from yadif. Can you provide samples including still graphics that are improved with this patch?
Hi Thomas,
Sorry for this thread going unanswered for so long. With the Vulkan hwaccel stuff sorted out, I wanted to come back to this. I discussed more with Lynne and we're no longer convinced this change is obviously desirable. You are right about the jumping behaviour - and although there are samples (eg: the burosch ones on samples.ffmpeg.org) which look better with the change, I don't think that's something to over index on.
I don't recall agreeing that this is desirable. bwdif_vulkan still looks better than bwdif or bwdif_cuda (and it's got proper edge handling).
Could you please prove this claim with some samples? I don't have any suitable hardware for Vulkan. Just saw that bwdif_vulkan was added to master. Somehow I didn't see the commit in the cvs log.
participants (3)
-
Lynne -
Philip Langdale -
Thomas Mundt