File size: 25,772 Bytes
a3cb3a7
3a1265d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
73df34b
3a1265d
73df34b
3a1265d
 
73df34b
3a1265d
 
 
 
 
 
73df34b
 
 
3a1265d
 
73df34b
 
3a1265d
73df34b
3a1265d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a3cb3a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3a1265d
a3cb3a7
 
 
 
 
3a1265d
a3cb3a7
 
3a1265d
a3cb3a7
 
 
 
 
 
 
 
 
 
 
 
 
3a1265d
a3cb3a7
 
 
 
 
 
 
3a1265d
a3cb3a7
 
 
 
 
 
 
 
 
3a1265d
a3cb3a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3a1265d
a3cb3a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
73df34b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a3cb3a7
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3a1265d
 
73df34b
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
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
import re
import os
import dataclasses
from enum import auto, Enum
from typing import List, Tuple
from collections import defaultdict
from .constants import PART_ORDER, COCO_KEYPOINT_NAME

def read_hoi_file_2_dict(hoi_config):
    hoi_dict = {}
    with open(hoi_config, "r", encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            if not line or line.startswith("#"):
                continue
            nums, obj, action = line.split()  # split on whitespace
            hoi_dict[int(nums)] = [obj, action]      # use nums as int; remove int() if you want string keys
    return hoi_dict

def read_part_state_file_2_dict(part_state_config):
    d = defaultdict(list)
    with open(part_state_config, "r", encoding="utf-8") as f:
        for line in f:
            line = line.strip()
            if not line or line.startswith("#"):
                continue

            key, val = line.split(":", 1)   # split only on first ":"
            key = key.strip()
            val = val.strip()
            d[key].append(val)
    return d

@dataclasses.dataclass
class Conversation:
    def __init__(self, system='', data_path=''):
        super().__init__()       
        if system == '':
           self.system = f"""
           You are an AI assistant for first-pass long-form HICO annotation. You will be given an image that contains a main human subject.
           Task:
           Write a detailed long description of the visual evidence in the image that supports the subject's action, with an emphasis on human body parts, posture, spatial configuration, and interactions with objects.

           Hints:
           You may be given hints about (1) the action and (2) related objects and possible supporting body parts. You should use these hints as anchors, and you may add other relevant visible evidence you observe.

           Required Constraints:
           - Start with ONE sentence that summarizes the main action in natural language.
           - When you mention any keypoint or body part, you MUST use names ONLY from: {COCO_KEYPOINT_NAME}.
           - Do NOT invent body-part names outside these sets (no synonyms, no paraphrases).
           - If you are unsure which name applies, either omit the body-part mention or choose the closest valid name from the lists.
           - The description must be long and detailed enough to serve as a first-pass annotation for later refinement.
           - Include as many relevant supporting details as are visibly justified, especially about contact, pose, orientation, support, and object interaction.
           - Write your description in clear, natural sentences grounded in visible evidence.

           Optional Constraints :
           - Prefer a rich multi-sentence paragraph rather than a short caption.
           - Cover multiple cues when available, such as limb placement, body balance, joint bending, contact points, and relative position to the object.
           - Write naturally. Avoid repeating the same sentence pattern.
           - If both sides contribute differently, describe them separately.
           """
        else:
            self.system = system
         
        self.hoi_reference = read_hoi_file_2_dict(os.path.join(data_path, 'Configs/hico_hoi_list.txt'))
        self.part_state_reference = read_part_state_file_2_dict(os.path.join(data_path, 'Configs/Part_State_76.txt'))

    def _humanpart2word(self, action_labels):
        action_labels_in_words = []
        part_state_keys = list(self.part_state_reference.keys())
        for d in action_labels:
            human_part_id = d['human_part']
            part_state_id = d['partstate']

            part_name = PART_ORDER[human_part_id]
            for key in part_state_keys:
                if key in part_name:
                    states = self.part_state_reference[key]
                    part_state = states[part_state_id]
            action_labels_in_words.append([part_name, part_state])
        return action_labels_in_words

    def _actionid2word(self, hoi_id):
        obj, act = self.hoi_reference[hoi_id]
        return obj, act

    def get_prompt(self, meta):
        hoi_obj = meta['hoi_obj']
        
        hoi_id = hoi_obj['hoi_id']
        obj_in_word, act_in_word = self._actionid2word(hoi_id)
        action_labels = hoi_obj['action_labels']
        action_labels_in_words = self._humanpart2word(action_labels)

        prompt = f"""
               Given the image, describe the visual evidence (especially body parts) that supports the action.
               Hints: The action to support is [{act_in_word} with {obj_in_word}]. Possible visual evidence cues include: {action_labels_in_words}.
               Use these cues as guidance. Only mention cues you can actually see in the image.
        """
        return prompt
   
@dataclasses.dataclass
class Conversation_For_Clean_Descrption:
    def __init__(self, system='', data_path=''):
        super().__init__()       
        if system == '':
           self.system = f"""
        You are a strict verifier and editor for pose-grounded action descriptions.

        You will be given:
        - Ground-truth action label(s) (GT).
        - A candidate description (may be verbose or include irrelevant evidence).
        - A closed list of allowed keypoint/body-part names.
        - A replacement mapping (e.g., hand→wrist, foot→ankle).

        Rules:
        1) First, check whether the candidate’s stated action matches the GT action(s).
        2) Then rewrite the description into exactly 2–3 sentences:
        - The first sentence must state the GT action (not the candidate action if it differs).
        - Keep only evidence that supports the GT action; delete unrelated evidence.
        - If a joint is mentioned both sides ALWAYS write as "left_wrist and right_wrist", "left_hip and right_hip", "left_ankle and right_ankle", etc.
        - When mentioning body parts/keypoints, you MUST use only names from the allowed list (exact match).
        - Apply the replacement mapping strictly; never output disallowed synonyms like “hand/foot” if they map to allowed names.
        - Do not add new evidence; only keep/condense evidence already present in the candidate.
        - A MUST-KEEP hint: required (joint, part_action) items that must appear in the final description (joint names may need replacement).

        Output format (plain text only): The refined 2–3 sentence description.
        No other text.
           """
        else:
            self.system = system
         
        self.hoi_reference = read_hoi_file_2_dict(os.path.join(data_path, 'Configs/hico_hoi_list.txt'))
        self.part_state_reference = read_part_state_file_2_dict(os.path.join(data_path, 'Configs/Part_State_76.txt'))
    
    def _replace_part_names(self, text):
        REPL = {
                "hand": "wrist",
                "hands": "wrists",
                "foot": "ankle",
                "feet": "ankles",
            }
        pattern = re.compile(r"\b(" + "|".join(map(re.escape, REPL.keys())) + r")\b", re.IGNORECASE)
        def _sub(m):
            w = m.group(0)
            out = REPL[w.lower()]
            # keep capitalization if you care
            if w[0].isupper():
                out = out.capitalize()
            return out
        return pattern.sub(_sub, text)

    def _humanpart2word(self, action_labels):
        action_labels_in_words = []
        part_state_keys = list(self.part_state_reference.keys())
        for d in action_labels:
            human_part_id = d['human_part']
            part_state_id = d['partstate']

            part_name = PART_ORDER[human_part_id]
            for key in part_state_keys:
                if key in part_name:
                    states = self.part_state_reference[key]
                    part_state = states[part_state_id]

            part_name = self._replace_part_names(part_name)
            action_labels_in_words.append([part_name, part_state])
        return action_labels_in_words

    def _actionid2word(self, hoi_id):
        obj, act = self.hoi_reference[hoi_id]
        return obj, act

    def get_prompt(self, meta):
        hoi_id = meta['hoi_id']
        obj_in_word, act_in_word = self._actionid2word(hoi_id)
        action_labels = meta['action_labels']
        action_labels_in_words = self._humanpart2word(action_labels)

        description = meta['description']
        description = self._replace_part_names(description)

        prompt = f"""
               GT action(s): {act_in_word, obj_in_word}
               Allowed keypoint names:
               {COCO_KEYPOINT_NAME}
               Replacement mapping:
               "hand" to "wrist", "foot" to "ankle"
               Candidate description:
               {description}
               Must-KEEP Hint:
               {action_labels_in_words}
                Please follow the system rules and output in the required plain-text format.
        """
        return prompt

@dataclasses.dataclass
class Conversation_For_Clean_Evidence:
    def __init__(self, system='', data_path=''):
        super().__init__()       
        if system == '':
           self.system = f"""
            You rewrite descriptions into NATURAL LANGUAGE evidence-only text.

            Output rules:
            - Write 2–3 complete sentences in natural English.
            - Do NOT mention the action or the subject (no "person", "he", "she", "they", "main", etc.).
            - Only describe evidence involving body parts/keypoints and part-level motions/contacts.
            - Every sentence must include at least one keypoint name from the allowed list (exact match).
            - Only use keypoint names from the allowed list; no other body-part words.
            - Never use generic joints (e.g., "wrist", "hip", "ankle") alone; If both sides are mentioned, use "left_wrist and right_wrist", "left_hip and right_hip", etc.
            - Apply the replacement mapping first (hand→wrist, foot→ankle, etc.), then enforce left/right by writing both sides.
            - Keep only evidence supported by the candidate; do not add new details.

            Style variety requirement:
            - Write like a reasoning use normal grammar, not lists, not "keypoint: ...".
            - Do not use the same starter phrase or the same connector in both sentences.
            - Example reasoning patterns (Can invent your own, but use different pattern):
            A) “With <keypoints/evidence>, <interpretation>.”  (no “suggesting/indicating”)
            B) “<Interpretation>; evidence includes <keypoints/evidence>.”  (semicolon style)
            C) “This is supported by <keypoints/evidence>, which <effect/constraint>.”  (“supported by” style)
            D) “Notably, <keypoints/evidence>; this points to <interpretation>.”  (“notably/points to” style)
            E) “<Keypoints/evidence> form(s) <configuration>, consistent with <interpretation>.” (“configuration” style)
           """
        else:
            self.system = system
         
        self.hoi_reference = read_hoi_file_2_dict(os.path.join(data_path, 'Configs/hico_hoi_list.txt'))
        self.part_state_reference = read_part_state_file_2_dict(os.path.join(data_path, 'Configs/Part_State_76.txt'))
    
    def _replace_part_names(self, text):
        REPL = {
                "hand": "wrist",
                "hands": "wrists",
                "foot": "ankle",
                "feet": "ankles",
            }
        pattern = re.compile(r"\b(" + "|".join(map(re.escape, REPL.keys())) + r")\b", re.IGNORECASE)
        def _sub(m):
            w = m.group(0)
            out = REPL[w.lower()]
            # keep capitalization if you care
            if w[0].isupper():
                out = out.capitalize()
            return out
        return pattern.sub(_sub, text)

    def _humanpart2word(self, action_labels):
        action_labels_in_words = []
        part_state_keys = list(self.part_state_reference.keys())
        for d in action_labels:
            human_part_id = d['human_part']
            part_state_id = d['partstate']

            part_name = PART_ORDER[human_part_id]
            for key in part_state_keys:
                if key in part_name:
                    states = self.part_state_reference[key]
                    part_state = states[part_state_id]

            part_name = self._replace_part_names(part_name)
            action_labels_in_words.append([part_name, part_state])
        return action_labels_in_words

    def _actionid2word(self, hoi_id):
        obj, act = self.hoi_reference[hoi_id]
        return obj, act

    def get_prompt(self, meta):
        hoi_id = meta['hoi_id']
        obj_in_word, act_in_word = self._actionid2word(hoi_id)
        action_labels = meta['action_labels']
        action_labels_in_words = self._humanpart2word(action_labels)

        description = meta['short_description']
        description = self._replace_part_names(description)
        prompt = f"""
               GT action(s): {act_in_word, obj_in_word}
               Allowed keypoint names:
               {COCO_KEYPOINT_NAME}
               Replacement mapping:
               "hand" to "wrist", "foot" to "ankle"
               Candidate description:
               {description}
               Must-KEEP Hint:
               {action_labels_in_words}
                Please follow the system rules and output in the required plain-text format.
        """
        return prompt

@dataclasses.dataclass
class Conversation_examiner:
    def __init__(self, system='', data_path=''):
        super().__init__()
        if system == '':
           self.system = f"""
           You are a strict checker and final editor for HICO action annotations.

           You will be given:
           - The ground-truth HICO action hint as [VERB, OBJECT].
           - Part-state hints derived from annotation labels.
           - One or more candidate texts, such as a long description, a short description, or an evidence-only description.

           Your task:
           - Judge whether the candidate texts are consistent with the target action.
           - Check whether the descriptions are grounded in plausible visible body-part evidence.
           - Check whether any mentioned body parts use valid COCO keypoint names only: {COCO_KEYPOINT_NAME}.
           - Detect unsupported claims, contradictions, object/action mismatches, left/right mistakes, and hallucinated joints or interactions.
           - Produce a final checked description after resolving any issues you can fix from the provided candidates and hints.

           Important checking rules:
           - The target action is defined by the provided HICO hint, not by the candidate text.
           - If a candidate text conflicts with the target action, fix the final checked description so it aligns with the target action.
           - If a candidate text includes body-part terms outside the allowed keypoint list, replace them with valid names when possible and record the issue.
           - If evidence is too vague, missing, or unrelated to the target action, remove unsupported content from the final checked description and record the issue.
           - Pay special attention to left/right consistency. If the candidate confuses left and right, or assigns evidence to the wrong side, correct it when the correct side is supported by the provided candidates and hints; otherwise remove the uncertain side-specific claim and record the issue.
           - Do not keep any joint claim that is not visible, not inferable from the provided evidence, or appears hallucinated. If a joint or body-part interaction cannot be supported, remove it and record the issue.
           - Do not invent new visual evidence that is not supported by the provided candidates and hints.
           - The final checked description should be concise, natural, and reliable.
           - Prefer the strongest grounded evidence among the provided candidates.
           - When side-specific evidence is uncertain, prefer a conservative description over a risky one.

           Output format:
           Return plain text in exactly this structure.

           Verdict: PASS or REVISED
           Action alignment: one short sentence
           Evidence grounding: one short sentence
           Keypoint-name validity: one short sentence
           Checked description:
           <final checked description>
           Issues:
           - item 1
           - item 2

           If there are no issues, write:
           Issues:
           - None
           """
        else:
            self.system = system

        self.hoi_reference = read_hoi_file_2_dict(os.path.join(data_path, 'Configs/hico_hoi_list.txt'))
        self.part_state_reference = read_part_state_file_2_dict(os.path.join(data_path, 'Configs/Part_State_76.txt'))

    def _replace_part_names(self, text):
        REPL = {
                "hand": "wrist",
                "hands": "wrists",
                "foot": "ankle",
                "feet": "ankles",
            }
        pattern = re.compile(r"\b(" + "|".join(map(re.escape, REPL.keys())) + r")\b", re.IGNORECASE)
        def _sub(m):
            w = m.group(0)
            out = REPL[w.lower()]
            if w[0].isupper():
                out = out.capitalize()
            return out
        return pattern.sub(_sub, text)

    def _humanpart2word(self, action_labels):
        action_labels_in_words = []
        part_state_keys = list(self.part_state_reference.keys())
        for d in action_labels:
            human_part_id = d['human_part']
            part_state_id = d['partstate']

            part_name = PART_ORDER[human_part_id]
            for key in part_state_keys:
                if key in part_name:
                    states = self.part_state_reference[key]
                    part_state = states[part_state_id]

            part_name = self._replace_part_names(part_name)
            action_labels_in_words.append([part_name, part_state])
        return action_labels_in_words

    def _actionid2word(self, hoi_id):
        obj, act = self.hoi_reference[hoi_id]
        return obj, act

    def get_prompt(self, meta):
        hoi_id = meta['hoi_id']
        obj_in_word, act_in_word = self._actionid2word(hoi_id)
        action_labels = meta['action_labels']
        action_labels_in_words = self._humanpart2word(action_labels)

        long_description = self._replace_part_names(meta.get('description', ''))
        refined_description = self._replace_part_names(meta.get('refined_description', ''))
        short_description = self._replace_part_names(meta.get('short_description', ''))
        action_description = self._replace_part_names(meta.get('action_description', ''))
        evidence_description = self._replace_part_names(meta.get('evidence_description', ''))

        prompt = f"""
               Target action hint: [{act_in_word}, {obj_in_word}]
               Part-state hints:
               {action_labels_in_words}

               Candidate long description:
               {long_description if long_description else '[Missing]'}

               Candidate refined description:
               {refined_description if refined_description else '[Missing]'}

               Candidate short description:
               {short_description if short_description else '[Missing]'}

               Candidate action description:
               {action_description if action_description else '[Missing]'}

               Candidate evidence description:
               {evidence_description if evidence_description else '[Missing]'}

               Check the candidates against the target action and part-state hints, produce the final checked description, and then follow the required output format exactly.
        """
        return prompt

@dataclasses.dataclass
class Conversation_For_Action_Pharse:
    def __init__(self, system='', data_path=''):
        super().__init__()       
        if system == '':
           self.system = f"""
           You are a visual captioning assistant.
           Given an image and an action hint in the form [VERB, OBJECT], output exactly one short English sentence describing that action in the image.

           Rules:
                •	Use only the provided VERB and OBJECT (you may adjust grammar: holds/holding; a/the; plural if needed).
                •	Output one sentence only.
                •	No extra details (no location, colors, emotions, reasons, scene context).
                •	No punctuation beyond the final period.
                •	If the subject is a person, use “The person” (not “man/woman/boy/girl”).
                •	If the action is not visible, still output a best-effort sentence using the hint.
           """
        else:
            self.system = system
         
        self.hoi_reference = read_hoi_file_2_dict(os.path.join(data_path, 'Configs/hico_hoi_list.txt'))
        self.part_state_reference = read_part_state_file_2_dict(os.path.join(data_path, 'Configs/Part_State_76.txt'))
    
    def _replace_part_names(self, text):
        REPL = {
                "hand": "wrist",
                "hands": "wrists",
                "foot": "ankle",
                "feet": "ankles",
            }
        pattern = re.compile(r"\b(" + "|".join(map(re.escape, REPL.keys())) + r")\b", re.IGNORECASE)
        def _sub(m):
            w = m.group(0)
            out = REPL[w.lower()]
            # keep capitalization if you care
            if w[0].isupper():
                out = out.capitalize()
            return out
        return pattern.sub(_sub, text)

    def _humanpart2word(self, action_labels):
        action_labels_in_words = []
        part_state_keys = list(self.part_state_reference.keys())
        for d in action_labels:
            human_part_id = d['human_part']
            part_state_id = d['partstate']

            part_name = PART_ORDER[human_part_id]
            for key in part_state_keys:
                if key in part_name:
                    states = self.part_state_reference[key]
                    part_state = states[part_state_id]

            part_name = self._replace_part_names(part_name)
            action_labels_in_words.append([part_name, part_state])
        return action_labels_in_words

    def _actionid2word(self, hoi_id):
        obj, act = self.hoi_reference[hoi_id]
        return obj, act

    def get_prompt(self, meta):
        hoi_id = meta['hoi_id']
        obj_in_word, act_in_word = self._actionid2word(hoi_id)
        action_labels = meta['action_labels']
        action_labels_in_words = self._humanpart2word(action_labels)

        description = meta['short_description']
        description = self._replace_part_names(description)
        prompt = f"""
               Hints: {act_in_word, obj_in_word}
               Write exactly one short sentence that follows the rules.
        """
        return prompt
    
@dataclasses.dataclass
class Conversation_For_COCO_Long_Description:
    def __init__(self, system='', data_path=''):
        super().__init__()       
        if system == '':
           self.system = f"""
           You are an AI assistant. You will be given an image that contains a main human subject.
           Task:
           Describe the visual evidence in the image that supports the subject’s action, with an emphasis on human body parts and their interactions with objects.

           Hints:
           You may be given hints about (1) the action and (2) related objects and possible supporting body parts. You can use these hints, but you may also add other relevant evidence you observe.

           Required Constraints:
           - Start with ONE sentence that summarizes the main action in natural language.
           - When you mention any keypoint or body part, you MUST use names ONLY from: {COCO_KEYPOINT_NAME}.
           - Do NOT invent body-part names outside these sets (no synonyms, no paraphrases).
           - If you are unsure which name applies, either omit the body-part mention or choose the closest valid name from the lists.
           - Write your description in clear, concise sentences grounded in visible evidence.

           Optional Constraints :
           - Write naturally. Avoid repeating the same sentence pattern.
           - Keep each evidence item to one line. No redundant "both left/right do the same" unless necessary.
           """
        else:
            self.system = system
         
        self.hoi_reference = read_hoi_file_2_dict(os.path.join(data_path, 'Configs/hico_hoi_list.txt'))
        self.part_state_reference = read_part_state_file_2_dict(os.path.join(data_path, 'Configs/Part_State_76.txt'))
       
    def _replace_part_names(self, text):
        REPL = {
                "hand": "wrist",
                "hands": "wrists",
                "foot": "ankle",
                "feet": "ankles",
            }
        pattern = re.compile(r"\b(" + "|".join(map(re.escape, REPL.keys())) + r")\b", re.IGNORECASE)
        def _sub(m):
            w = m.group(0)
            out = REPL[w.lower()]
            # keep capitalization if you care
            if w[0].isupper():
                out = out.capitalize()
            return out
        return pattern.sub(_sub, text)

    def get_prompt(self, meta):
        
        prompt = f"""
               Hint: you may consider use the actions in the below dictionary {self.part_state_reference}
               Given the image, describe the visual evidence (especially body parts) that supports the action.
        """
        return prompt


if __name__ == "__main__":
   pass