Detect soft 2:3 pulldown (telecine) files and decode them at 23.976.

DVD rips from NTSC DVDs are sometimes (always?) encoded using
soft 2:3 pulldown.  The video frames are actually 23.976 but
FFmpeg detects them as 29.97.  With the current approach of the video
decoder ignoring most PTSs and assuming a constant frame rate
it is vital that the file contains the number of frames per second
that the detected frame rate predicts.

This fixes large sync errors with NTSC DVD rips (#1790).

3 files changed, 48 insertions, 5 deletions

diff --git a/src/lib/ffmpeg_content.cc b/src/lib/ffmpeg_content.cc
index f3c0d01cb..1fb95535f 100644
--- a/src/lib/ffmpeg_content.cc
+++ b/src/lib/ffmpeg_content.cc
@@ -337,6 +337,14 @@ FFmpegContent::examine (shared_ptr<const Film> film, shared_ptr<Job> job)
 		set_default_colour_conversion ();
 	}
 
+	if (examiner->has_video() && examiner->pulldown() && video_frame_rate() && fabs(*video_frame_rate() - 29.97) < 0.001) {
+		/* FFmpeg has detected this file as 29.97 and the examiner thinks it is using "soft" 2:3 pulldown (telecine).
+		 * This means we can treat it as a 23.976fps file.
+		 */
+		set_video_frame_rate (24000.0 / 1001);
+		video->set_length (video->length() * 24.0 / 30);
+	}
+
 #ifdef DCPOMATIC_VARIANT_SWAROOP
 	_id = examiner->id ();
 #endif
diff --git a/src/lib/ffmpeg_examiner.cc b/src/lib/ffmpeg_examiner.cc
index cc6a62ff6..87bb8ce95 100644
--- a/src/lib/ffmpeg_examiner.cc
+++ b/src/lib/ffmpeg_examiner.cc
@@ -18,6 +18,7 @@
 
 */
 
+#include "dcpomatic_log.h"
 #include "ffmpeg_examiner.h"
 #include "ffmpeg_content.h"
 #include "job.h"
@@ -43,16 +44,24 @@ DCPOMATIC_ENABLE_WARNINGS
 using std::string;
 using std::cout;
 using std::max;
+using std::vector;
 using boost::shared_ptr;
 using boost::optional;
 using namespace dcpomatic;
 
 
+/* This is how many frames from the start of any video that we will examine to see if we
+ * can spot soft 2:3 pull-down ("telecine").
+ */
+static const int PULLDOWN_CHECK_FRAMES = 16;
+
+
 /** @param job job that the examiner is operating in, or 0 */
 FFmpegExaminer::FFmpegExaminer (shared_ptr<const FFmpegContent> c, shared_ptr<Job> job)
 	: FFmpeg (c)
 	, _video_length (0)
 	, _need_video_length (false)
+	, _pulldown (false)
 {
 	/* Find audio and subtitle streams */
 
@@ -106,9 +115,15 @@ DCPOMATIC_DISABLE_WARNINGS
 	/* Run through until we find:
 	 *   - the first video.
 	 *   - the first audio for each stream.
+	 *   - the top-field-first and repeat-first-frame values ("temporal_reference") for the first PULLDOWN_CHECK_FRAMES video frames.
 	 */
 
 	int64_t const len = _file_group.length ();
+	/* A string which we build up to describe the top-field-first and repeat-first-frame values for the first few frames.
+	 * It would be nicer to use something like vector<bool> here but we want to search the array for a pattern later,
+	 * and a string seems a reasonably neat way to do that.
+	 */
+	string temporal_reference;
 	while (true) {
 		int r = av_read_frame (_format_context, &_packet);
 		if (r < 0) {
@@ -127,7 +142,7 @@ DCPOMATIC_DISABLE_WARNINGS
 DCPOMATIC_ENABLE_WARNINGS
 
 		if (_video_stream && _packet.stream_index == _video_stream.get()) {
-			video_packet (context);
+			video_packet (context, temporal_reference);
 		}
 
 		bool got_all_audio = true;
@@ -143,7 +158,7 @@ DCPOMATIC_ENABLE_WARNINGS
 
 		av_packet_unref (&_packet);
 
-		if (_first_video && got_all_audio) {
+		if (_first_video && got_all_audio && temporal_reference.size() >= PULLDOWN_CHECK_FRAMES) {
 			/* All done */
 			break;
 		}
@@ -173,6 +188,13 @@ DCPOMATIC_ENABLE_WARNINGS
 		DCPOMATIC_ASSERT (fabs (*_rotation - 90 * round (*_rotation / 90)) < 2);
 	}
 
+	LOG_GENERAL("Temporal reference was %1", temporal_reference);
+	if (temporal_reference.find("T2T3B2B3T2T3B2B3") || temporal_reference.find("B2B3T2T3B2B3T2T3")) {
+		/* The magical sequence (taken from mediainfo) suggests that 2:3 pull-down is in use */
+		_pulldown = true;
+		LOG_GENERAL_NC("Suggest that this may be 2:3 pull-down (soft telecine)");
+	}
+
 #ifdef DCPOMATIC_VARIANT_SWAROOP
 	AVDictionaryEntry* e = av_dict_get (_format_context->metadata, SWAROOP_ID_TAG, 0, 0);
 	if (e) {
@@ -182,12 +204,16 @@ DCPOMATIC_ENABLE_WARNINGS
 }
 
 
+/** @param temporal_reference A string to which we should add two characters per frame;
+ *  the first   is T or B depending on whether it's top- or bottom-field first,
+ *  ths seconds is 3 or 2 depending on whether "repeat_pict" is true or not.
+ */
 void
-FFmpegExaminer::video_packet (AVCodecContext* context)
+FFmpegExaminer::video_packet (AVCodecContext* context, string& temporal_reference)
 {
 	DCPOMATIC_ASSERT (_video_stream);
 
-	if (_first_video && !_need_video_length) {
+	if (_first_video && !_need_video_length && temporal_reference.size() >= PULLDOWN_CHECK_FRAMES) {
 		return;
 	}
 
@@ -203,6 +229,10 @@ DCPOMATIC_ENABLE_WARNINGS
 				_format_context->streams[_video_stream.get()]
 				).get_value_or (ContentTime ()).frames_round (video_frame_rate().get ());
 		}
+		if (temporal_reference.size() < PULLDOWN_CHECK_FRAMES) {
+			temporal_reference += (_frame->top_field_first ? "T" : "B");
+			temporal_reference += (_frame->repeat_pict ? "3" : "2");
+		}
 	}
 }
 
diff --git a/src/lib/ffmpeg_examiner.h b/src/lib/ffmpeg_examiner.h
index 9ffb8421b..4f5110ca0 100644
--- a/src/lib/ffmpeg_examiner.h
+++ b/src/lib/ffmpeg_examiner.h
@@ -77,6 +77,10 @@ public:
 		return _rotation;
 	}
 
+	bool pulldown () const {
+		return _pulldown;
+	}
+
 #ifdef DCPOMATIC_VARIANT_SWAROOP
 	boost::optional<std::string> id () const {
 		return _id;
@@ -84,7 +88,7 @@ public:
 #endif
 
 private:
-	void video_packet (AVCodecContext *);
+	void video_packet (AVCodecContext *, std::string& temporal_reference);
 	void audio_packet (AVCodecContext *, boost::shared_ptr<FFmpegAudioStream>);
 
 	std::string stream_name (AVStream* s) const;
@@ -101,6 +105,7 @@ private:
 	bool _need_video_length;
 
 	boost::optional<double> _rotation;
+	bool _pulldown;
 
 #ifdef DCPOMATIC_VARIANT_SWAROOP
 	boost::optional<std::string> _id;