android_cast/AV_QUALITY_QA_SESSION.md

AV quality enhancement — Q&A session notes

Research and recommendations for movie cast vs video cast quality enhancement in the Android Cast project. Combines user requirements, comparison tables, algorithm choices, placement (sender vs receiver), and fit with the current codebase.

Status: Planning / research only — not implemented.
Related code (today): sender ScreenCastService → VideoEncoder / AudioEncoder → network; receiver network → VideoDecoder / LibvpxCapableVideoDecoder / AudioDecoder → TextureView / AudioTrack.

---

1. User requirements (preliminary conditions)

1. Content is either movie cast or video cast (live screen/camera); both must be supported with different tuning.
2. Both video and audio need quality enhancement.
3. Two processing shapes:
   • 2.1 Temporal (N=2 frames/windows): small gap between sequential video frames → predict/enhance using deltas between frames; same idea for audio (overlapping buffers).
   • 2.2 Single frame/chunk: static image from sink/source; one video frame; one audio PCM chunk.
4. 2.3 Reassembly: output stream must keep original PTS/DTS (or fixed delay with uniform compensation); avoid changing frame/sample count on live paths unless explicitly designed.

Comparison dimensions requested: implementation details, difficulty, realtime on modern ARM (GPU Android/Linux where applicable), input vs output quality, single vs 2-frame vs stream (with delay), multithreading.

---

2. Mode overview

Mode	Video input	Audio input	Typical delay	Best when
2.2 Single	One buffer / frame	10–40 ms PCM chunk	0–1 frame (~0–33 ms @ 30 fps)	Live video cast, UI, games
2.1 Two-frame (N=2)	Pair with small temporal gap	Overlapping windows	~1 frame (+ overlap)	Motion; mild temporal help
Stream (3–8)	Short FIFO	Ring + STFT state	100–400 ms+	Movie cast, VOD-like
Heavy stream ML	Long context	Streaming neural	200 ms–2 s+	Offline / non-live only

Quality (typical): stream/temporal > 2-frame > single frame
Latency (best first): single frame > 2-frame > stream

Policy split:

Cast type	Video	Audio
Video cast (live)	2.2 default; optional light 2.1 (+1 frame max)	2.2 (RNNoise + limiter)
Movie cast	2.1 or 3–4 frame TNR / MCTF-lite	2-window spectral or DeepFilterNet2 streaming

---

3. Video — method comparison

Method	Implementation	Difficulty	ARM @ 720p30	ARM @ 1080p30	GPU (Android/Linux)	Quality vs input	PTS/DTS	Multithread
Single: bilateral / guided filter denoise	GLES/Vulkan on YUV; NEON fallback	Low	Yes	Yes (tuned)	High	Small–medium	Trivial	Yes (tiles)
Single: tone / CLAHE-lite	LUT + luma histogram tiles	Low–med	Yes	Yes	Medium	Medium in flat scenes	Trivial	Yes
Single: light SR (ESPCN-class tiny CNN)	TFLite/NCNN → NNAPI/GPU	Medium	Maybe	Tight	High	Medium (edges/text)	Trivial	Yes (async infer)
Single: heavy SR (Real-ESRGAN, SwinIR)	Large CNN	High	No	No	BW-bound	High but slow	Trivial	Limited
2-frame: DIS / Farneback flow + warp blend	Flow on downscaled luma → warp prev	Medium	Yes	Maybe	Med–high	Medium	+1 frame	Yes
2-frame: residual / delta enhance	enhanced = f(curr) + α·(curr − warped_prev)	Medium	Yes	Maybe	Medium	Med–high on repetitive motion	+1 frame	Yes
2-frame: RIFE-lite / FILM-small	Interpolate between frames	Med–high	Borderline 720p	Unlikely 1080p30	High if custom	High for large gaps	Timing care	Yes
Stream: IIR TNR / motion-mask accumulate	3–4 frame buffer	Medium	Yes	Maybe	High	High static; trail risk	+2–3 frames	Yes
Stream: ML temporal (VRT/RVRT-class)	Stateful temporal CNN	Very high	No live	No	Medium	Very high	Buffer delay	Some

---

4. Audio — method comparison

Method	Implementation	Difficulty	ARM realtime 48 kHz	GPU	Quality vs input	PTS/DTS	Multithread
Single: DC block + EQ + limiter	Biquads per chunk	Low	Yes	Low	Small	Preserve samples	Yes
Single: RNNoise / DeepFilterNet (10 ms)	TFLite/ONNX streaming	Medium	Yes mono; maybe stereo	NNAPI	High for noise	+10–20 ms	Yes (infer thread)
2-window: Wiener / min-stats STFT	20 ms, 50% hop	Medium	Yes	Low	Med–high	+1 hop (~10 ms)	Yes
2-window: PLC	Predict from last 2 frames	Low–med	Yes	Low	High only on loss	Keep clock	Yes
Stream: multiband + AGC	Ring buffer envelopes	Low–med	Yes	Low	Medium intelligibility	5–30 ms	Yes
Stream: DeepFilterNet2	Stateful STFT+GRU	Med–high	Yes modern ARM	NNAPI	High	20–40 ms	Yes

---

5. Single vs 2-frame vs stream — decision matrix

Criterion	Single (2.2)	2-frame (2.1)	Stream (3–8)
Implementation complexity	Lowest	Medium	Highest
Encoder integration	Easiest	+1 frame pipeline	Delay queue
ARM realtime 720p30	Best	Good	Good (TNR); poor heavy ML
ARM realtime 1080p30	Good	Marginal	Marginal
GPU fit	Best	Good (warp)	Good (TNR)
Quality gain (video)	Low–medium	Medium + motion	High static / medium action
Quality gain (audio)	Medium	Med–high	High steady noise
Artifact risk	Banding, oversharpen	Ghosting, warp error	Trails, smear
PTS/DTS	Trivial	Fixed +1 frame delay	Uniform delay all outputs
Multithread	Capture ∥ GPU ∥ encode	Flow ∥ enhance	Buffer thread + pool
Movie cast	OK quick polish	Recommended	Recommended if delay OK
Video cast	Recommended	Optional 1-frame	Not recommended live

---

6. PTS/DTS reassembly (2.3)

Approach	Description	Modes
In-place payload replace	Same PTS/DTS, duration; swap enhanced payload	Single
Delay compensation	Queue enhanced frames; release with original PTS	2-frame, stream
Composition timestamp	presentationTimeUs from source → enhanced buffer (MediaCodec)	All on Android
Audio	Keep sample count constant → PTS unchanged	Prefer for live

Rule: Do not change frame/sample count on live cast unless explicit insert/drop policy exists.

---

7. Pipeline placement: sender (2.1) vs receiver (2.2)

User question

• 2.1 Sender: enhancement between capture (camera/screen) and encoder.
• 2.2 Receiver: enhancement between decoded stream and renderer.
• User intuition: 2.1 has limited value because encoder lossy output destroys pre-enhance detail; increases sender load.

Assessment

Claim	Verdict
Encoder “kills” sharpen/SR/local contrast pushed pre-encode	True
2.1 is pointless	False for denoise-before-encode (helps bitrate/artifacts at same rate)
2.1 burns sender CPU for little gain	Often true on thermally limited phones

Preference (consensus from session)

Placement	Role
Receiver (2.2)	Primary perceived quality — deblock, denoise, mild sharpen; optional 2-frame for movie
Sender (2.1)	Optional compression prep only (bilateral/NLMeans denoise), not main beauty pipeline
Both	Receiver polishes; sender light or off on live video cast

Current app hooks:

Pipeline	Integration point	Difficulty (1=easy, 5=hard)
2.2 Video receiver	After VideoDecoder / LibvpxCapableVideoDecoder, before TextureView	3–4 (new GLES: decode → FBO → display)
2.2 Audio receiver	AudioDecoder PCM before AudioTrack	2
2.1 Video sender	Between capture and VideoEncoder.prepare() Surface	4–5 (intermediate EGL Surface, rotation)
2.1 Audio sender	Before AudioEncoder	2–3
2-frame video	Ring buffer +1 frame delay	+1 on above

Recommended implementation fork: decoded-buffer → GLES → Surface (consistent for MediaCodec and libvpx VP9), not only TextureView overlay.

---

8. Named algorithm pipelines (best practices)

A. Receiver — live video cast (default)

Video (2.2, GPU-first):

1. Post-decode deblocking filter (lite H.264-style / mild dering; VP9 loop filter already in decode)
2. Guided filter or bilateral filter (edge-preserving denoise) on luma
3. Mild unsharp mask (limit on blocky regions)
4. Optional: gamma / BT.1886 + CLAHE-lite

Audio (2.2):

1. RNNoise (10 ms frames)
2. Limiter + high-pass (DC removal)
3. Optional: SpeexDSP preprocess (AGC, mild denoise)

B. Receiver — movie cast (+1 frame delay)

Video (2.1):

1. DIS or Farneback optical flow (downscaled)
2. MCTF-lite or IIR temporal denoise on stable regions
3. Spatial chain from A (guided/bilateral + mild unsharp)

Audio (2.1):

1. STFT overlap-add + Wiener filter, or DeepFilterNet2 (streaming)
2. Multiband compressor

C. Sender — optional compression prep only

Video: Bilateral or fast NLMeans only when bitrate-starved — no Real-ESRGAN / strong sharpen pre-encode.

Audio: SpeexDSP preprocess or light RNNoise before AAC.

D. Avoid for realtime in this project

Algorithm	Reason
Real-ESRGAN, SwinIR, BasicVSR++	Too heavy ARM @ 720p–1080p30 live
RIFE / FILM @ 1080p30 full rate	Borderline; movie-only, lower rate
BM3D, VRT/RVRT	Offline-tier
Heavy 2.1 sender SR before VP9	Encoder erases; thermals

---

9. Ranked preferences (quality benefit vs implementation in android cast)

Descending order — build in this sequence:

Rank	Choice	Placement	Core algorithms	Quality benefit	Impl difficulty (this app)
1	Receiver video 2.2 (GPU)	2.2	Guided/bilateral + deblock + mild unsharp	High vs blocky stream	3–4
2	Receiver audio 2.2	2.2	RNNoise + limiter (+ SpeexDSP)	High for noise	2
3	Receiver video 2.1 (movie)	2.2 + 1f delay	DIS/Farneback + MCTF-lite / IIR TNR + spatial	Higher static/grain	4
4	Receiver audio 2.1 (movie)	2.2	DeepFilterNet2 or Wiener STFT	Medium–high	3
5	Sender audio light prep	2.1	SpeexDSP / RNNoise → AAC	Low–medium	2–3
6	Sender video denoise only	2.1	Bilateral / fast NLMeans (not SR)	Low–medium (bitrate)	4–5
7	Sender video 2.1 beauty	2.1	Flow + enhance pre-encode	Low (washed by codec)	5
8	Heavy SR either side	either	Real-ESRGAN class	High offline only	5, not realtime

---

10. Multithreading pipeline (target architecture)

[Capture / Network decode] → queue → [Enhance worker(s)] → queue → [Encode / Render]
                                    ↑
                              N=2: previous frame
                              Stream: ring buffer

Stage	Threads	Notes
Capture / decode callback	1	Non-blocking
Enhance (GPU)	1–2	Double-buffered FBO / AHardwareBuffer
CPU fallback (NEON)	2–4	Tile-based
ML audio	1	Async from video
Encode / display	1	MediaCodec async; TextureView main

Golden rule: max one frame enhancement queue for video cast; 2–4 frames for movie cast.

Receiver already has videoDecodeThread in ReceiverCastService — natural host for GL enhance stage.

---

11. Short answers to numbered conditions

#	Requirement	Preference
1	Movie or video cast	Different latency policies; same codebase, different presets
2	Process both A+V	Shared audio chain; video policy switches by cast mode
2.1.1	2-frame video deltas	DIS/Farneback + residual — not full RIFE on ARM live
2.1.2	2-frame audio	Overlap-add STFT or streaming DeepFilterNet
2.2.1	Single image video	GPU bilateral/guided + tone; tiny SR @ 720p only if ever
2.2.2	Single chunk audio	RNNoise + limiter
2.3	PTS/DTS reassembly	In-place or fixed-delay queue; no timeline shift on live

---

12. Bottom line

Goal	Best option
Lowest latency + ARM-safe	Single frame receiver (2.2) + GPU shaders
Best quality per ms delay	2-frame receiver (2.1) + 2-window/stream audio
Best absolute quality (movie)	Small stream buffer (3–4) video TNR + streaming audio ML
Avoid realtime cast	Heavy per-frame SR, RIFE @ 1080p30, deep temporal video CNN
Placement	Receiver-first; sender denoise-only at most

User’s encoder observation is correct for “beauty” enhancement pre-encode; denoise-before-encode remains the only strong sender-side exception.

---

13. Other work in same project session (context only)

Unrelated to AV enhancement but done in parallel on this branch:

• OTA optional .otabundle.zip, crash reporter (:crashwatcher), HEADER.txt automation, PHP crash backend — see docs/OTA.md, docs/CRASH_REPORTER.md, examples/crash_reporter/.

---

14. Open decisions when implementation starts

1. Enhance on display path (TextureView) vs decode buffer → GLES → Surface (recommended: latter).
2. User-toggle vs automatic preset for movie vs video cast.
3. Whether sender denoise-only preset is worth thermals on low-end devices.
4. GPU API: OpenGL ES 3.x compute vs Vulkan (ES3 often faster to integrate in Android media apps).

---

Document generated from AV quality Q&A session. Update as prototypes prove FPS/latency on target devices.