File size: 5,516 Bytes
5029d29
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
"""Frame-push interpreter for the web app (browser-camera architecture).

Unlike `CameraInterpreter` (which owns a server-side webcam — only usable
locally), this receives frames pushed from the browser over HTTP, runs the
holistic tracker + recognizer, and returns compact landmarks + glosses for the
browser to draw. This is what makes the app deployable to Hugging Face, where
the server has no camera.

Tuned for CPU-only hosts (HF free tier): MediaPipe at `model_complexity=0`,
face refinement off, server-side drawing skipped. MediaPipe's graph is stateful
and not thread-safe, so all processing is serialised behind one lock — fine for
a demo; for many concurrent users you'd shard sessions.
"""

from __future__ import annotations

import threading
import time
from collections import deque
from typing import Deque, List, Optional

import numpy as np

from .camera import InterpreterPipeline
from .landmarks import points_payload
from .recognizer import Recognition


class FrameSession:
    """Process-wide singleton that turns pushed frames into landmarks + glosses."""

    _instance: Optional["FrameSession"] = None
    _init_lock = threading.Lock()

    def __init__(
        self,
        *,
        templates_path: str = "signs/lsf_signs.json",
        model_path: Optional[str] = None,
        model_complexity: int = 0,
        refine_face_landmarks: bool = False,
        max_transcript: int = 200,
    ) -> None:
        self._pipeline = InterpreterPipeline(
            templates_path=templates_path,
            model_path=model_path,
            model_complexity=model_complexity,
            refine_face_landmarks=refine_face_landmarks,
        )
        self._lock = threading.Lock()
        self._transcript: Deque[dict] = deque(maxlen=max_transcript)
        self._seq = 0
        self._fps = 0.0
        self._last_t = 0.0
        # Depth capability reported by the client device (dual-pixel detection).
        self._depth_source = "mediapipe"
        self._depth_reason = "No client depth reported yet."

    @classmethod
    def instance(cls, **kwargs) -> "FrameSession":
        with cls._init_lock:
            if cls._instance is None:
                cls._instance = cls(**kwargs)
            return cls._instance

    # -- depth capability (set by the browser) -------------------------------
    def set_depth_capability(self, *, dual_pixel: bool, reason: str = "", label: str = "") -> None:
        if dual_pixel:
            self._depth_source = "dual_pixel"
            self._depth_reason = f"Dual-pixel/depth-capable device: {label or 'detected'}."
        else:
            self._depth_source = "mediapipe"
            self._depth_reason = reason or "No dual-pixel depth on this device; using MediaPipe z."

    # -- main entry point -----------------------------------------------------
    def process_jpeg(self, jpeg_bytes: bytes) -> dict:
        """Decode a pushed JPEG, run detection + recognition, return a payload."""
        import cv2  # local import keeps module import cheap

        arr = np.frombuffer(jpeg_bytes, dtype=np.uint8)
        frame = cv2.imdecode(arr, cv2.IMREAD_COLOR)
        if frame is None:
            return {"ok": False, "error": "Could not decode frame."}

        with self._lock:
            result = self._pipeline.process(frame, draw=False)
            reco = result.recognition
            if reco is not None:
                self._append(reco)
            now = time.time()
            if self._last_t:
                self._fps = 0.85 * self._fps + 0.15 * (1.0 / max(now - self._last_t, 1e-3))
            self._last_t = now

        return {
            "ok": True,
            "landmarks": points_payload(result.landmarks),
            "reco": self._reco_dict(reco) if reco else None,
            "status": self.status,
        }

    # -- transcript -----------------------------------------------------------
    def _append(self, reco: Recognition) -> None:
        self._seq += 1
        self._transcript.append(self._reco_dict(reco, seq=self._seq))

    @staticmethod
    def _reco_dict(reco: Recognition, seq: int = 0) -> dict:
        return {
            "seq": seq,
            "gloss": reco.gloss,
            "confidence": round(reco.confidence, 3),
            "source": reco.source,
            "t": reco.timestamp,
        }

    def transcript(self, since: int = 0) -> List[dict]:
        with self._lock:
            return [e for e in self._transcript if e["seq"] > since]

    def clear_transcript(self) -> None:
        with self._lock:
            self._transcript.clear()

    # -- recording / vocabulary proxies --------------------------------------
    def start_recording(self, label: str) -> None:
        self._pipeline.recognizer.start_recording(label)

    def stop_recording(self) -> Optional[str]:
        return self._pipeline.recognizer.stop_recording()

    def cancel_recording(self) -> None:
        self._pipeline.recognizer.cancel_recording()

    def template_counts(self) -> dict:
        return self._pipeline.recognizer.template_counts()

    def delete_gloss(self, gloss: str) -> bool:
        return self._pipeline.recognizer.delete_gloss(gloss)

    @property
    def status(self) -> dict:
        rec = self._pipeline.recognizer
        return {
            "depth_source": self._depth_source,
            "depth_reason": self._depth_reason,
            "fps": round(self._fps, 1),
            "recording": rec.is_recording,
            "trained": len(rec.vocabulary),
        }