File size: 15,972 Bytes
a83d3e5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
# -*- coding: utf-8 -*-
"""GROVE 定性结果图: 每case = 视频帧时间线(证据帧红框) + Q气泡 + GT行 + Ours行

拆成两张图:
  qualitative_cases_qa.pdf      = OVO-Bench (3) + MMLifelong (1) + ESTP (2)   —— 问答类
  qualitative_cases_service.pdf = EgoServe 即时安全 (1) + 中长期例程优化 (1)  —— 主动服务类
同时仍输出合并版 qualitative_cases.pdf (向后兼容)。

每个 case 用 `strip` 字段指定自己的帧目录; 帧文件名 = {prefix}_{k}.jpg, 证据帧后缀 _EV.jpg。
"""
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from matplotlib.patches import FancyBboxPatch, Rectangle
import matplotlib.image as mpimg
import numpy as np, os, textwrap

FC = "/mnt/tokyo-ai/gst/paper_final_results/figure_cases"
S_OVO  = f"{FC}/ovo_backward/strip"          # OVO + EgoServe 即时(stairs)
S_MM   = f"{FC}/mmlifelong/strip"
S_ESTP = f"{FC}/estp/strip"
S_LT   = f"{FC}/egoserve_longterm/strip"
OUTDIR = "/mnt/tokyo-ai/gst/paper_final_results/figure_data"
os.makedirs(OUTDIR, exist_ok=True)

# 配色
C_FRAME_BG = "#EAF2FB"   # 帧排底色 浅蓝
C_Q        = "#DDF2F3"   # Q气泡 浅青
C_Q_EDGE   = "#7FBFC4"
C_GT       = "#F0F0F0"   # GT 浅灰
C_GT_EDGE  = "#B8B8B8"
C_OURS     = "#DFF3E0"   # Ours 浅绿
C_OURS_EDGE= "#5FA968"
C_EV       = "#D62728"   # 证据帧红框
C_HL       = "#1F5FA8"   # 问题关键短语蓝
C_OK       = "#2E7D32"

# ===========================================================================
# 所有 case 均为「真实答对」样本, Q / GT / 我们的回答全部取自评测文件英文原文
# ===========================================================================

CASES_QA = [
 dict(
   tag="OVO-Bench · Backward (EPM)",
   strip=S_OVO, prefix="microwave", n=6,
   span="evidence @ 03:54   →   question @ 07:47   (look-back  3.9 min)",
   ev_cap="closes the microwave door",
   q_pre="", q_hl="Did I close the microwave?", q_post="",
   opts="A. yes    B. no",
   gt="A. yes",
   ours="A. yes",
   ours_note="“Video facts show the camera-wearer pressing down and closing the silver microwave door (DAY1-00:07:44) and then pressing the control buttons to set 1:00, so the microwave was closed.”",
 ),
 dict(
   tag="OVO-Bench · Backward (EPM)",
   strip=S_OVO, prefix="scraper", n=6,
   span="evidence @ 00:25–00:30   →   question @ 19:58   (look-back  19.6 min)",
   ev_cap="the dough cutter in use",
   q_pre="", q_hl="Where was the dough cutter before I picked it?", q_post="",
   opts="A. Table    B. Floor    C. Cabinet    D. Shelf",
   gt="A. Table",
   ours="Table",
   ours_note="“The entity log shows a ‘dough scraper [metal blade, wooden handle]’ located on the table just before the pick-up event, so it was on the table.”",
 ),
 dict(
   tag="OVO-Bench · Backward (ASI)",
   strip=S_OVO, prefix="pierogi", n=6,
   span="anchor step “boil” @ 07:30   →   successor step @ 08:20   (spans the full 9-min recipe)",
   ev_cap="pan-frying the boiled pierogies",
   q_pre="What does the person do after ", q_hl="boil the wrapped pierogies in boiling salted water",  q_post="",
   opts="A. fry the boiled pierogies with a little oil in a pan and then serve    B. combine the meat with half of the onion salt bay leaf dried\nmushrooms peppercorns and water    C. wrap the meat mixture in the small dough portions    D. keep rolling until you get 1 to 2 mm thick dough",
   gt="A. fry the boiled pierogies with a little oil in a pan and then serve",
   ours="A. fry the boiled pierogies with a little oil in a pan and then serve",
   ours_note="“The video shows golden-brown pan-fried dumplings and the sequence after boiling the sealed pierogies proceeds to pan-fry them to crisp the bottoms, matching option A.”",
 ),
 # ---------------- MMLifelong: 25h+ 超长视频里的几秒细节 ----------------
 dict(
   tag="MM-Lifelong · month scale  (Entity Recognition)",
   strip=S_MM, prefix="hotpot", n=6,
   span="evidence @ 03:40:15 of a 5.4-h video  —  hour 61 of a 99-hour, 22-video multi-day recording   (needle-in-a-haystack)",
   ev_cap="sipping the green soy-milk bottle",
   q_pre="", q_hl="What drink did the streamer have when eating hot pot in Chengdu, China?", q_post="",
   opts="",
   gt="Soy milk",
   ours="Soy milk",
   ours_note="“During the Chengdu hot pot segment, the facts show him picking up and sipping a green bottle labeled ‘Soy Milk’ through a straw while eating.”   —   official judge (gpt-5) Final Score = 5/5",
 ),
 # ---------------- ESTP-Bench: 半主动服务 ----------------
 dict(
   tag="ESTP-Bench · Contextual QA  (Object Relative Context)",
   strip=S_ESTP, prefix="stopsign", n=6,
   span="user asks @ 31:49  →  system stays silent and keeps watching  →  fires @ 31:58, the moment the sign enters view   (54-min drive)",
   ev_cap="the red STOP sign comes into view",
   q_pre="", q_hl="Could you remind me when I see the red road sign saying “STOP”?", q_post="",
   opts="",
   gt="There is a red road sign on your right hand side saying “STOP”.",
   ours="You see the red ‘STOP’ sign now.",
   ours_note="emitted at t = 1918.0 s, inside the GT evidence window 1916.0–1920.5 s   —   DeepSeek judge correctness = 5.0 / 5",
 ),
 dict(
   tag="ESTP-Bench · Contextual QA  (Object Relative Context)",
   strip=S_ESTP, prefix="ladder", n=6,
   span="question @ 14:05   →   answered @ 14:11 from the streaming memory   (22-min studio session)",
   ev_cap="red ladder against the white wall",
   q_pre="", q_hl="Could you tell me where the red ladder in the room is?", q_post="",
   opts="",
   gt="The red ladder stands by the wall on the left.",
   ours="The red ladder is standing against the plain white wall to your left.",
   ours_note="emitted at t = 851.0 s, inside the GT evidence window 851.0–872.5 s   —   DeepSeek judge correctness = 5.0 / 5",
 ),
]

CASES_SERVICE = [
 dict(
   tag="EgoServe / EgoLife · Proactive Service   —   Instant  (safety)",
   strip=S_OVO, prefix="stairs", n=6,
   span="continuous egocentric stream   →   hazard trigger @ DAY3 12:12:00  (no user query — the system speaks first)",
   ev_cap="descending stairs, eyes on phone",
   q_pre="", q_hl="[ no question asked ]", q_post="  user is walking down the stairs while reading messages on the phone",
   opts="",
   gt="Be careful; walking down the stairs while staring at your phone is risky.",
   ours="Careful — you're looking at your phone while going down the stairs, which could\nmake you lose your footing. Please pause and watch your step.",
   ours_note="matched to GT event safety_001 (sub-type “safety”, overlapping time window)",
 ),
 # ---------------- 中长期: 跨时段 pattern 层 ----------------
 dict(
   tag="EgoServe / EgoLife · Proactive Service   —   Long-Term  (routine_optimization)",
   strip=S_LT, prefix="assembly", n=6,
   span="the same furniture-assembly routine recurs on DAY1 12:52  and  DAY4 12:25  →  recognised again and acted on @ DAY4 17:13  (a 4-day, cross-session pattern — not a single-clip cue)",
   ev_cap="assembling the storage drawers again",
   q_pre="", q_hl="[ no question asked ]", q_post="  user starts assembling yet another storage unit",
   opts="",
   gt="I see you're assembling another storage unit. Would you like me to search for an\nassembly video for this model to help you out?",
   ours="You're assembling storage drawers again. Want me to pull up your last assembly\nphotos/checklist, auto-display the instruction steps for this model, or create a\nquick parts/tools checklist to streamline the process?",
   ours_note="“ROUTINE_CANDIDATES: multiple prior furniture-assembly episodes (DAY4 12:25:30, DAY1 12:52:00, etc.). Current window shows assembly restarting around 17:13:00.”   —   strict sub-type match, GT DAY4 17:13:01 vs ours DAY4 17:13:00. Requires the cross-session pattern layer: the trigger is the repetition itself, which no single frame contains.",
 ),
]

# ---- 版面度量 (figure 坐标, 0-1) ----
L, R = 0.035, 0.965
W = R - L

H_TAG   = 0.34
H_STRIP = 1.52
H_SPAN  = 0.24
H_Q     = 0.50
H_OPT   = 0.36
H_GT    = 0.40
H_OURS  = 0.86
H_GAP   = 0.50


def case_height(c):
    h = H_TAG + H_STRIP + H_SPAN + H_Q + H_GT + H_OURS
    if c["opts"]:
        h += H_OPT * (1 + c["opts"].count("\n"))
    h += 0.30 * c["ours"].count("\n")
    h += 0.30 * c["gt"].count("\n")
    h += 0.16 * (len(textwrap.fill(c["ours_note"], 118).split("\n")) - 1)
    return h


def render(cases, title, outname):
    FIG_W = 13.6
    FIG_H = sum(case_height(c) for c in cases) + H_GAP * (len(cases) - 1) + 1.05

    fig = plt.figure(figsize=(FIG_W, FIG_H))
    fig.patch.set_facecolor("white")

    def fw(x_in): return x_in / FIG_W
    def fh(y_in): return y_in / FIG_H

    def rbox(x, y, w, h, fc, ec, lw=1.1, r=0.012, z=2):
        p = FancyBboxPatch((x, y), w, h, boxstyle=f"round,pad=0,rounding_size={r}",
                           facecolor=fc, edgecolor=ec, linewidth=lw,
                           transform=fig.transFigure, zorder=z,
                           mutation_aspect=FIG_W / FIG_H)
        fig.patches.append(p); return p

    def txt(x, y, s, size=9.5, color="#1a1a1a", weight="normal", ha="left", va="center",
            style="normal", family="DejaVu Sans", z=5, **kw):
        return fig.text(x, y, s, fontsize=size, color=color, fontweight=weight, ha=ha, va=va,
                        style=style, family=family, zorder=z, **kw)

    fig.text(0.5, 1 - fh(0.30), title, ha="center", va="center",
             fontsize=15.5, fontweight="bold", color="#12253d")

    y_cur = 1.0 - fh(0.58)

    for c in cases:
        S = c["strip"]

        # ---------- tag ----------
        y_cur -= fh(H_TAG)
        txt(L, y_cur + fh(0.185), c["tag"], size=10.5, weight="bold", color="#12253d")
        bw, bh = fw(1.02), fh(0.20)
        rbox(R - bw, y_cur + fh(0.085), bw, bh, C_OURS, C_OURS_EDGE, lw=1.0, r=0.010)
        txt(R - bw / 2, y_cur + fh(0.085) + bh / 2, "✓  CORRECT", size=8.6, weight="bold",
            color=C_OK, ha="center")
        fig.add_artist(plt.Line2D([L, R], [y_cur + fh(0.045)] * 2, color="#c9d6e4", lw=0.9,
                                  transform=fig.transFigure, zorder=1))

        # ---------- 帧排 ----------
        y_cur -= fh(H_STRIP)
        rbox(L, y_cur, W, fh(H_STRIP), C_FRAME_BG, "#BBD3EA", lw=1.0, r=0.010, z=1)

        n = c["n"]
        pad = fw(0.13)
        gap = fw(0.075)
        avail = W - 2 * pad - gap * (n - 1)
        cw = avail / n
        inner_h = fh(H_STRIP) - fh(0.62)
        y_img = y_cur + fh(0.30)

        for k in range(n):
            fn = f"{S}/{c['prefix']}_{k}.jpg"
            fn_ev = f"{S}/{c['prefix']}_{k}_EV.jpg"
            is_ev = os.path.exists(fn_ev)
            path = fn_ev if is_ev else fn
            if not os.path.exists(path):
                continue
            im = mpimg.imread(path)
            ih, iw = im.shape[:2]
            ar = iw / ih
            h_ax = inner_h
            w_ax = h_ax * ar * (FIG_H / FIG_W)
            if w_ax > cw:
                w_ax = cw
                h_ax = w_ax / ar * (FIG_W / FIG_H)
            x_ax = L + pad + k * (cw + gap) + (cw - w_ax) / 2
            yy = y_img + (inner_h - h_ax) / 2
            ax = fig.add_axes([x_ax, yy, w_ax, h_ax], zorder=3)
            ax.imshow(im); ax.set_xticks([]); ax.set_yticks([])
            for sp in ax.spines.values():
                sp.set_edgecolor(C_EV if is_ev else "#9DB4CC")
                sp.set_linewidth(3.0 if is_ev else 0.8)
            if is_ev:
                fig.patches.append(Rectangle((x_ax - fw(0.028), yy - fh(0.028)),
                     w_ax + fw(0.056), h_ax + fh(0.056), fill=False, edgecolor=C_EV,
                     linewidth=1.5, transform=fig.transFigure, zorder=4))
                txt(x_ax + w_ax / 2, yy + h_ax + fh(0.105), "key evidence frame", size=8.0,
                    weight="bold", color=C_EV, ha="center")
                txt(x_ax + w_ax / 2, yy - fh(0.115), c["ev_cap"], size=7.6, color=C_EV,
                    ha="center", style="italic")
            else:
                txt(x_ax + w_ax / 2, yy - fh(0.115), "·", size=7.0, color="#9DB4CC", ha="center")

        txt(L + fw(0.10), y_cur + fh(0.13), "video timeline  (uniformly sampled frames)",
            size=8.0, color="#5B7FA6", style="italic")

        # ---------- 跨度说明 ----------
        y_cur -= fh(H_SPAN)
        txt(L + fw(0.04), y_cur + fh(0.10), c["span"], size=8.6, color="#7a5a00", style="italic")

        # ---------- Q 气泡 ----------
        y_cur -= fh(H_Q)
        rbox(L, y_cur + fh(0.05), W, fh(H_Q) - fh(0.05), C_Q, C_Q_EDGE, lw=1.2, r=0.011)
        qy = y_cur + fh(0.05) + (fh(H_Q) - fh(0.05)) / 2
        txt(L + fw(0.16), qy, "Q:", size=11, weight="bold", color="#14606a")
        xoff = L + fw(0.46)
        rend = fig.canvas.get_renderer()
        for seg, col, wt in [(c["q_pre"], "#1a1a1a", "normal"),
                             (c["q_hl"], C_HL, "bold"),
                             (c["q_post"], "#1a1a1a", "normal")]:
            if not seg:
                continue
            t = txt(xoff, qy, seg, size=10.2, color=col, weight=wt)
            bb = t.get_window_extent(renderer=rend)
            xoff += bb.width / (FIG_W * fig.dpi)

        # ---------- options ----------
        if c["opts"]:
            nl = 1 + c["opts"].count("\n")
            y_cur -= fh(H_OPT * nl)
            txt(L + fw(0.62), y_cur + fh(H_OPT * nl) / 2, c["opts"], size=8.5,
                color="#4a5b6b", va="center")

        # ---------- GT ----------
        gt_lines = 1 + c["gt"].count("\n")
        gt_in = H_GT + 0.24 * (gt_lines - 1)
        y_cur -= fh(gt_in)
        rbox(L, y_cur + fh(0.04), W, fh(gt_in) - fh(0.06), C_GT, C_GT_EDGE, lw=1.0, r=0.011)
        gy = y_cur + fh(0.04) + (fh(gt_in) - fh(0.06)) / 2
        txt(L + fw(0.16), gy, "GT", size=9.8, weight="bold", color="#4a4a4a")
        txt(L + fw(1.42), gy, c["gt"], size=9.8, color="#2b2b2b", linespacing=1.35)

        # ---------- Ours ----------
        ans_lines = 1 + c["ours"].count("\n")
        note = textwrap.fill(c["ours_note"], 118)
        note_lines = 1 + note.count("\n")
        box_in = 0.30 + 0.185 * ans_lines + 0.145 * note_lines
        y_cur -= fh(box_in)
        bh2 = fh(box_in) - fh(0.08)
        rbox(L, y_cur + fh(0.04), W, bh2, C_OURS, C_OURS_EDGE, lw=1.4, r=0.011)
        oy_top = y_cur + fh(0.04) + bh2
        ny = oy_top - fh(0.18)
        txt(L + fw(0.16), ny, "GROVE", size=9.8, weight="bold", color=C_OK, va="top")
        txt(L + fw(0.72), ny, "(ours)", size=8.4, color="#5c8a63", style="italic", va="top")
        txt(L + fw(1.42), ny, c["ours"], size=9.8, color="#14421a", weight="bold",
            va="top", linespacing=1.35)
        txt(L + fw(1.42), ny - fh(0.185 * ans_lines + 0.06), note, size=7.6, color="#4d6b52",
            style="italic", va="top", linespacing=1.4)

        y_cur -= fh(H_GAP)

    fig.text(0.5, fh(0.16),
      "Red box = the single frame containing the answer evidence; all other frames are uniformly sampled distractors that do not support the answer.",
      ha="center", va="center", fontsize=8.3, color="#6b7a8a", style="italic")

    fig.savefig(f"{OUTDIR}/{outname}.pdf", format="pdf", bbox_inches="tight", facecolor="white")
    fig.savefig(f"{OUTDIR}/{outname}.png", dpi=170, bbox_inches="tight", facecolor="white")
    plt.close(fig)
    print(f"saved -> {OUTDIR}/{outname}.pdf   ({len(cases)} cases, {FIG_H:.1f} in tall)")


render(CASES_QA,
       "Qualitative results of GROVE on long-horizon egocentric question answering",
       "qualitative_cases_qa")

render(CASES_SERVICE,
       "Qualitative results of GROVE on proactive service (instant hazard vs. long-term routine)",
       "qualitative_cases_service")

# 合并版(向后兼容)
render(CASES_QA + CASES_SERVICE,
       "Qualitative results of GROVE on long-horizon egocentric benchmarks",
       "qualitative_cases")