Skip to content

zamba.models.densepose.densepose_manager

Attributes

DENSEPOSE_AVAILABLE = True module-attribute

MODELS = dict(animals=dict(config=str(Path(__file__).parent / 'assets' / 'densepose_rcnn_R_50_FPN_soft_animals_I0_finetune_16k.yaml'), densepose_weights_url='https://dl.fbaipublicfiles.com/densepose/cse/densepose_rcnn_R_50_FPN_soft_animals_I0_finetune_16k/270727112/model_final_421d28.pkl', weights='zamba_densepose_model_final_421d28.pkl', viz_class=DensePoseOutputsVertexVisualizer, viz_class_kwargs=dict()), chimps=dict(config=str(Path(__file__).parent / 'assets' / 'densepose_rcnn_R_50_FPN_soft_chimps_finetune_4k.yaml'), densepose_weights_url='https://dl.fbaipublicfiles.com/densepose/cse/densepose_rcnn_R_50_FPN_soft_chimps_finetune_4k/253146869/model_final_52f649.pkl', weights='zamba_densepose_model_final_52f649.pkl', viz_class=DensePoseOutputsTextureVisualizer, viz_class_kwargs=dict(texture_atlases_dict={'chimp_5029': get_texture_atlas(str(Path(__file__).parent / 'assets' / 'chimp_texture_colors_flipped.tif'))}), anatomy_color_mapping=str(Path(__file__).parent / 'assets' / 'chimp_5029_parts.csv'))) module-attribute

Classes

DensePoseManager

Source code in zamba/models/densepose/densepose_manager.py
 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
class DensePoseManager:
    def __init__(
        self,
        model=MODELS["chimps"],
        model_cache_dir: Path = Path(".zamba_cache"),
        download_region=RegionEnum("us"),
    ):
        """Create a DensePoseManager object.

        Parameters
        ----------
        model : dict, optional (default MODELS['chimps'])
            A dictionary with the densepose model defintion like those defined in MODELS.
        """
        if not DENSEPOSE_AVAILABLE:
            raise ImportError(
                "Densepose not installed. See: https://zamba.drivendata.org/docs/stable/models/densepose/#installation"
            )

        # setup configuration for densepose
        self.cfg = get_cfg()
        add_densepose_config(self.cfg)

        self.cfg.merge_from_file(model["config"])

        if not (model_cache_dir / model["weights"]).exists():
            model_cache_dir.mkdir(parents=True, exist_ok=True)
            self.cfg.MODEL.WEIGHTS = download_weights(
                model["weights"], model_cache_dir, download_region
            )

        # automatically use CPU if no cuda available
        if not torch.cuda.is_available():
            self.cfg.MODEL.DEVICE = "cpu"

        self.cfg.freeze()

        logging.getLogger("fvcore").setLevel("CRITICAL")  # silence noisy detectron2 logging
        # set up predictor with the configuration
        self.predictor = DefaultPredictor(self.cfg)

        # we have a specific texture atlas for chimps with relevant regions
        # labeled that we can use instead of the default segmentation
        self.visualizer = model["viz_class"](
            self.cfg,
            device=self.cfg.MODEL.DEVICE,
            **model.get("viz_class_kwargs", {}),
        )

        # set up utilities for use with visualizer
        self.vis_extractor = create_extractor(self.visualizer)
        self.vis_embedder = build_densepose_embedder(self.cfg)
        self.vis_class_to_mesh_name = get_class_to_mesh_name_mapping(self.cfg)
        self.vis_mesh_vertex_embeddings = {
            mesh_name: self.vis_embedder(mesh_name).to(self.cfg.MODEL.DEVICE)
            for mesh_name in self.vis_class_to_mesh_name.values()
            if self.vis_embedder.has_embeddings(mesh_name)
        }

        if "anatomy_color_mapping" in model:
            self.anatomy_color_mapping = pd.read_csv(model["anatomy_color_mapping"], index_col=0)
        else:
            self.anatomy_color_mapping = None

    def predict_image(self, image):
        """Run inference to get the densepose results for an image.

        Parameters
        ----------
        image :
            numpy array (unit8) of an image in BGR format or path to an image

        Returns
        -------
        tuple
            Returns the image array as passed or loaded and the the densepose Instances as results.
        """
        if isinstance(image, (str, Path)):
            image = read_image(image, format="BGR")

        return image, self.predict(image)

    def predict_video(self, video, video_loader_config=None, pbar=True):
        """Run inference to get the densepose results for a video.

        Parameters
        ----------
        video :
            numpy array (uint8) of a a video in BGR layout with time dimension first or path to a video
        video_loader_config : VideoLoaderConfig, optional
            A video loader config for loading videos (uses all defaults except pix_fmt="bgr24")
        pbar : bool, optional
            Whether to display a progress bar, by default True

        Returns
        -------
        tuple
            Tuple of (video_array, list of densepose results per frame)
        """
        if isinstance(video, (str, Path)):
            video = load_video_frames(video, config=video_loader_config)

        pbar = tqdm if pbar else lambda x, **kwargs: x

        return video, [
            self.predict_image(img)[1] for img in pbar(video, desc="Frames")
        ]  # just the predictions

    def predict(self, image_arr):
        """Main call to DensePose for inference. Runs inference on an image array.

        Parameters
        ----------
        image_arr : numpy array
            BGR image array

        Returns
        -------
        Instances
            Detection instances with boxes, scores, and densepose estimates.
        """
        with torch.no_grad():
            instances = self.predictor(image_arr)["instances"]

        return instances

    def serialize_video_output(self, instances, filename=None, write_embeddings=False):
        serialized = {
            "frames": [
                self.serialize_image_output(
                    frame_instances, filename=None, write_embeddings=write_embeddings
                )
                for frame_instances in instances
            ]
        }

        if filename is not None:
            with Path(filename).open("w") as f:
                json.dump(serialized, f, indent=2)

        return serialized

    def serialize_image_output(self, instances, filename=None, write_embeddings=False):
        """Convert the densepose output into Python-native objects that can
            be written and read with json.

        Parameters
        ----------
        instances : Instance
            The output from the densepose model
        filename : (str, Path), optional
            If not None, the filename to write the output to, by default None
        """
        if isinstance(instances, list):
            img_height, img_width = instances[0].image_size
        else:
            img_height, img_width = instances.image_size

        boxes = instances.get("pred_boxes").tensor
        scores = instances.get("scores").tolist()
        labels = instances.get("pred_classes").tolist()

        try:
            pose_result = instances.get("pred_densepose")
        except KeyError:
            pose_result = None

        # include embeddings + segmentation if they exist and they are requested
        write_embeddings = write_embeddings and (pose_result is not None)

        serialized = {
            "instances": [
                {
                    "img_height": img_height,
                    "img_width": img_width,
                    "box": boxes[i].cpu().tolist(),
                    "score": scores[i],
                    "label": {
                        "value": labels[i],
                        "mesh_name": self.vis_class_to_mesh_name[labels[i]],
                    },
                    "embedding": pose_result.embedding[[i], ...].cpu().tolist()
                    if write_embeddings
                    else None,
                    "segmentation": pose_result.coarse_segm[[i], ...].cpu().tolist()
                    if write_embeddings
                    else None,
                }
                for i in range(len(instances))
            ]
        }

        if filename is not None:
            with Path(filename).open("w") as f:
                json.dump(serialized, f, indent=2)

        return serialized

    def deserialize_output(self, instances_dict=None, filename=None):
        if filename is not None:
            with Path(filename).open("r") as f:
                instances_dict = json.load(f)

        # handle image case
        is_image = False
        if "frames" not in instances_dict:
            instances_dict = {"frames": [instances_dict]}
            is_image = True

        frames = []
        for frame in instances_dict["frames"]:
            heights, widths, boxes, scores, labels, embeddings, segmentations = zip(
                *[
                    (
                        i["img_height"],
                        i["img_width"],
                        i["box"],
                        i["score"],
                        i["label"]["value"],
                        i["embedding"] if i["embedding"] is not None else [np.nan],
                        i["segmentation"] if i["segmentation"] is not None else [np.nan],
                    )
                    for i in frame["instances"]
                ]
            )

            frames.append(
                Instances(
                    (heights[0], widths[0]),
                    pred_boxes=boxes,
                    scores=scores,
                    pred_classes=labels,
                    pred_densepose=DensePoseEmbeddingPredictorOutput(
                        embedding=torch.tensor(embeddings),
                        coarse_segm=torch.tensor(segmentations),
                    ),
                )
            )

        # if image or single frame, just return the instance
        if is_image:
            return frames[0]
        else:
            return frames

    def visualize_image(self, image_arr, outputs, output_path=None):
        """Visualize the pose information.

        Parameters
        ----------
        image_arr : numpy array (unit8) BGR
            The numpy array representing the image.
        outputs :
            The outputs from running DensePoseManager.predict*
        output_path : str or Path, optional
            If not None, write visualization to this path; by default None

        Returns
        -------
        numpy array (unit8) BGR
            DensePose outputs visualized on top of the image.
        """
        bw_image = cv2.cvtColor(image_arr, cv2.COLOR_BGR2GRAY)
        bw_image = np.tile(bw_image[:, :, np.newaxis], [1, 1, 3])
        data = self.vis_extractor(outputs)
        image_vis = self.visualizer.visualize(bw_image, data)

        if output_path is not None:
            cv2.imwrite(str(output_path), image_vis)

        return image_vis

    def anatomize_image(self, visualized_img_arr, outputs, output_path=None):
        """Convert the pose information into the percent of pixels in the detection
            bounding box that correspond to each part of the anatomy in an image.

        Parameters
        ----------
        visualized_img_arr : numpy array (unit8) BGR
            The numpy array the image after the texture has been visualized (by calling DensePoseManager.visualize_image).
        outputs :
            The outputs from running DensePoseManager.predict*

        Returns
        -------
        pandas.DataFrame
            DataFrame with percent of pixels of the bounding box that correspond to each anatomical part
        """
        if self.anatomy_color_mapping is None:
            raise ValueError(
                "No anatomy_color_mapping provided to track anatomy; did you mean to use a different MODEL?"
            )

        # no detections, return empty df for joining later (e.g., in anatomize_video)
        if not outputs:
            return pd.DataFrame([])

        _, _, N, bboxes_xywh, pred_classes = self.visualizer.extract_and_check_outputs_and_boxes(
            self.vis_extractor(outputs)
        )

        all_detections = []
        for n in range(N):
            x, y, w, h = bboxes_xywh[n].int().cpu().numpy()
            detection_area = visualized_img_arr[y : y + h, x : x + w]

            detection_stats = {
                name: (detection_area == np.array([[[color.B, color.G, color.R]]]))
                .all(axis=-1)
                .sum()
                / (h * w)  # calc percent of bounding box with this color
                for name, color in self.anatomy_color_mapping.iterrows()
            }

            detection_stats["x"] = x
            detection_stats["y"] = y
            detection_stats["h"] = h
            detection_stats["w"] = w

            all_detections.append(detection_stats)

        results = pd.DataFrame(all_detections)

        if output_path is not None:
            results.to_csv(output_path, index=False)

        return results

    def visualize_video(
        self, video_arr, outputs, output_path=None, frame_size=None, fps=30, pbar=True
    ):
        """Visualize the pose information on a video

        Parameters
        ----------
        video_arr : numpy array (unit8) BGR, time first
            The numpy array representing the video.
        outputs :
            The outputs from running DensePoseManager.predict*
        output_path : str or Path, optional
            If not None, write visualization to this path (should be .mp4); by default None
        frame_size : (innt, float), optional
            If frame_size is float, scale up or down by that float value; if frame_size is an integer,
            set width to that size and scale height appropriately.
        fps : int
            frames per second for output video if writing; defaults to 30
        pbar : bool
            display a progress bar

        Returns
        -------
        numpy array (unit8) BGR
            DensePose outputs visualized on top of the image.
        """
        pbar = tqdm if pbar else lambda x, **kwargs: x

        out_frames = np.array(
            [
                self.visualize_image(
                    image_arr,
                    output,
                )
                for image_arr, output in pbar(
                    zip(video_arr, outputs), total=video_arr.shape[0], desc="Visualize frames"
                )
            ]
        )

        if output_path is not None:
            # get new size for output video if scaling
            if frame_size is None:
                frame_size = video_arr.shape[2]  # default to same size

            # if float, scale as a multiple
            if isinstance(frame_size, float):
                frame_width = round(video_arr.shape[2] * frame_size)
                frame_height = round(video_arr.shape[1] * frame_size)

            # if int, use as width of the video and scale height proportionally
            elif isinstance(frame_size, int):
                frame_width = frame_size
                scale = frame_width / video_arr.shape[2]
                frame_height = round(video_arr.shape[1] * scale)

            # setup output for writing
            output_path = output_path.with_suffix(".mp4")
            out = cv2.VideoWriter(
                str(output_path),
                cv2.VideoWriter_fourcc(*"mp4v"),
                max(1, int(fps)),
                (frame_width, frame_height),
            )

            for f in pbar(out_frames, desc="Write frames"):
                if (f.shape[0] != frame_height) or (f.shape[1] != frame_width):
                    f = cv2.resize(
                        f,
                        (frame_width, frame_height),
                        # https://stackoverflow.com/a/51042104/1692709
                        interpolation=(
                            cv2.INTER_LINEAR if f.shape[1] < frame_width else cv2.INTER_AREA
                        ),
                    )
                out.write(f)

            out.release()

        return out_frames

    def anatomize_video(self, visualized_video_arr, outputs, output_path=None, fps=30):
        """Convert the pose information into the percent of pixels in the detection
            bounding box that correspond to each part of the anatomy in a video.

        Parameters
        ----------
        visualized_video_arr : numpy array (unit8) BGR
            The numpy array the video after the texture has been visualized (by calling DensePoseManager.visualize_video).
        outputs :
            The outputs from running DensePoseManager.predict*

        Returns
        -------
        numpy array (unit8) BGR
            DensePose outputs visualized on top of the image.
        """
        all_detections = []

        for ix in range(visualized_video_arr.shape[0]):
            detection_df = self.anatomize_image(visualized_video_arr[ix, ...], outputs[ix])
            detection_df["frame"] = ix
            detection_df["seconds"] = ix / fps
            all_detections.append(detection_df)

        results = pd.concat(all_detections)

        if output_path is not None:
            results.to_csv(output_path, index=False)

        return results

Attributes

anatomy_color_mapping = pd.read_csv(model['anatomy_color_mapping'], index_col=0) instance-attribute
cfg = get_cfg() instance-attribute
predictor = DefaultPredictor(self.cfg) instance-attribute
vis_class_to_mesh_name = get_class_to_mesh_name_mapping(self.cfg) instance-attribute
vis_embedder = build_densepose_embedder(self.cfg) instance-attribute
vis_extractor = create_extractor(self.visualizer) instance-attribute
vis_mesh_vertex_embeddings = {mesh_name: self.vis_embedder(mesh_name).to(self.cfg.MODEL.DEVICE) for mesh_name in self.vis_class_to_mesh_name.values() if self.vis_embedder.has_embeddings(mesh_name)} instance-attribute
visualizer = model['viz_class'](self.cfg, device=self.cfg.MODEL.DEVICE, None=model.get('viz_class_kwargs', {})) instance-attribute

Functions

__init__(model = MODELS['chimps'], model_cache_dir: Path = Path('.zamba_cache'), download_region = RegionEnum('us'))

Create a DensePoseManager object.

Parameters
dict, optional (default MODELS['chimps'])

A dictionary with the densepose model defintion like those defined in MODELS.

Source code in zamba/models/densepose/densepose_manager.py
 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
def __init__(
    self,
    model=MODELS["chimps"],
    model_cache_dir: Path = Path(".zamba_cache"),
    download_region=RegionEnum("us"),
):
    """Create a DensePoseManager object.

    Parameters
    ----------
    model : dict, optional (default MODELS['chimps'])
        A dictionary with the densepose model defintion like those defined in MODELS.
    """
    if not DENSEPOSE_AVAILABLE:
        raise ImportError(
            "Densepose not installed. See: https://zamba.drivendata.org/docs/stable/models/densepose/#installation"
        )

    # setup configuration for densepose
    self.cfg = get_cfg()
    add_densepose_config(self.cfg)

    self.cfg.merge_from_file(model["config"])

    if not (model_cache_dir / model["weights"]).exists():
        model_cache_dir.mkdir(parents=True, exist_ok=True)
        self.cfg.MODEL.WEIGHTS = download_weights(
            model["weights"], model_cache_dir, download_region
        )

    # automatically use CPU if no cuda available
    if not torch.cuda.is_available():
        self.cfg.MODEL.DEVICE = "cpu"

    self.cfg.freeze()

    logging.getLogger("fvcore").setLevel("CRITICAL")  # silence noisy detectron2 logging
    # set up predictor with the configuration
    self.predictor = DefaultPredictor(self.cfg)

    # we have a specific texture atlas for chimps with relevant regions
    # labeled that we can use instead of the default segmentation
    self.visualizer = model["viz_class"](
        self.cfg,
        device=self.cfg.MODEL.DEVICE,
        **model.get("viz_class_kwargs", {}),
    )

    # set up utilities for use with visualizer
    self.vis_extractor = create_extractor(self.visualizer)
    self.vis_embedder = build_densepose_embedder(self.cfg)
    self.vis_class_to_mesh_name = get_class_to_mesh_name_mapping(self.cfg)
    self.vis_mesh_vertex_embeddings = {
        mesh_name: self.vis_embedder(mesh_name).to(self.cfg.MODEL.DEVICE)
        for mesh_name in self.vis_class_to_mesh_name.values()
        if self.vis_embedder.has_embeddings(mesh_name)
    }

    if "anatomy_color_mapping" in model:
        self.anatomy_color_mapping = pd.read_csv(model["anatomy_color_mapping"], index_col=0)
    else:
        self.anatomy_color_mapping = None
anatomize_image(visualized_img_arr, outputs, output_path = None)

Convert the pose information into the percent of pixels in the detection bounding box that correspond to each part of the anatomy in an image.

Parameters
numpy array (unit8) BGR

The numpy array the image after the texture has been visualized (by calling DensePoseManager.visualize_image).

outputs

The outputs from running DensePoseManager.predict*

Returns

pandas.DataFrame DataFrame with percent of pixels of the bounding box that correspond to each anatomical part

Source code in zamba/models/densepose/densepose_manager.py
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
def anatomize_image(self, visualized_img_arr, outputs, output_path=None):
    """Convert the pose information into the percent of pixels in the detection
        bounding box that correspond to each part of the anatomy in an image.

    Parameters
    ----------
    visualized_img_arr : numpy array (unit8) BGR
        The numpy array the image after the texture has been visualized (by calling DensePoseManager.visualize_image).
    outputs :
        The outputs from running DensePoseManager.predict*

    Returns
    -------
    pandas.DataFrame
        DataFrame with percent of pixels of the bounding box that correspond to each anatomical part
    """
    if self.anatomy_color_mapping is None:
        raise ValueError(
            "No anatomy_color_mapping provided to track anatomy; did you mean to use a different MODEL?"
        )

    # no detections, return empty df for joining later (e.g., in anatomize_video)
    if not outputs:
        return pd.DataFrame([])

    _, _, N, bboxes_xywh, pred_classes = self.visualizer.extract_and_check_outputs_and_boxes(
        self.vis_extractor(outputs)
    )

    all_detections = []
    for n in range(N):
        x, y, w, h = bboxes_xywh[n].int().cpu().numpy()
        detection_area = visualized_img_arr[y : y + h, x : x + w]

        detection_stats = {
            name: (detection_area == np.array([[[color.B, color.G, color.R]]]))
            .all(axis=-1)
            .sum()
            / (h * w)  # calc percent of bounding box with this color
            for name, color in self.anatomy_color_mapping.iterrows()
        }

        detection_stats["x"] = x
        detection_stats["y"] = y
        detection_stats["h"] = h
        detection_stats["w"] = w

        all_detections.append(detection_stats)

    results = pd.DataFrame(all_detections)

    if output_path is not None:
        results.to_csv(output_path, index=False)

    return results
anatomize_video(visualized_video_arr, outputs, output_path = None, fps = 30)

Convert the pose information into the percent of pixels in the detection bounding box that correspond to each part of the anatomy in a video.

Parameters
numpy array (unit8) BGR

The numpy array the video after the texture has been visualized (by calling DensePoseManager.visualize_video).

outputs

The outputs from running DensePoseManager.predict*

Returns

numpy array (unit8) BGR DensePose outputs visualized on top of the image.

Source code in zamba/models/densepose/densepose_manager.py
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
def anatomize_video(self, visualized_video_arr, outputs, output_path=None, fps=30):
    """Convert the pose information into the percent of pixels in the detection
        bounding box that correspond to each part of the anatomy in a video.

    Parameters
    ----------
    visualized_video_arr : numpy array (unit8) BGR
        The numpy array the video after the texture has been visualized (by calling DensePoseManager.visualize_video).
    outputs :
        The outputs from running DensePoseManager.predict*

    Returns
    -------
    numpy array (unit8) BGR
        DensePose outputs visualized on top of the image.
    """
    all_detections = []

    for ix in range(visualized_video_arr.shape[0]):
        detection_df = self.anatomize_image(visualized_video_arr[ix, ...], outputs[ix])
        detection_df["frame"] = ix
        detection_df["seconds"] = ix / fps
        all_detections.append(detection_df)

    results = pd.concat(all_detections)

    if output_path is not None:
        results.to_csv(output_path, index=False)

    return results
deserialize_output(instances_dict = None, filename = None)
Source code in zamba/models/densepose/densepose_manager.py
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
def deserialize_output(self, instances_dict=None, filename=None):
    if filename is not None:
        with Path(filename).open("r") as f:
            instances_dict = json.load(f)

    # handle image case
    is_image = False
    if "frames" not in instances_dict:
        instances_dict = {"frames": [instances_dict]}
        is_image = True

    frames = []
    for frame in instances_dict["frames"]:
        heights, widths, boxes, scores, labels, embeddings, segmentations = zip(
            *[
                (
                    i["img_height"],
                    i["img_width"],
                    i["box"],
                    i["score"],
                    i["label"]["value"],
                    i["embedding"] if i["embedding"] is not None else [np.nan],
                    i["segmentation"] if i["segmentation"] is not None else [np.nan],
                )
                for i in frame["instances"]
            ]
        )

        frames.append(
            Instances(
                (heights[0], widths[0]),
                pred_boxes=boxes,
                scores=scores,
                pred_classes=labels,
                pred_densepose=DensePoseEmbeddingPredictorOutput(
                    embedding=torch.tensor(embeddings),
                    coarse_segm=torch.tensor(segmentations),
                ),
            )
        )

    # if image or single frame, just return the instance
    if is_image:
        return frames[0]
    else:
        return frames
predict(image_arr)

Main call to DensePose for inference. Runs inference on an image array.

Parameters
numpy array

BGR image array

Returns

Instances Detection instances with boxes, scores, and densepose estimates.

Source code in zamba/models/densepose/densepose_manager.py
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
def predict(self, image_arr):
    """Main call to DensePose for inference. Runs inference on an image array.

    Parameters
    ----------
    image_arr : numpy array
        BGR image array

    Returns
    -------
    Instances
        Detection instances with boxes, scores, and densepose estimates.
    """
    with torch.no_grad():
        instances = self.predictor(image_arr)["instances"]

    return instances
predict_image(image)

Run inference to get the densepose results for an image.

Parameters
image

numpy array (unit8) of an image in BGR format or path to an image

Returns

tuple Returns the image array as passed or loaded and the the densepose Instances as results.

Source code in zamba/models/densepose/densepose_manager.py
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
def predict_image(self, image):
    """Run inference to get the densepose results for an image.

    Parameters
    ----------
    image :
        numpy array (unit8) of an image in BGR format or path to an image

    Returns
    -------
    tuple
        Returns the image array as passed or loaded and the the densepose Instances as results.
    """
    if isinstance(image, (str, Path)):
        image = read_image(image, format="BGR")

    return image, self.predict(image)
predict_video(video, video_loader_config = None, pbar = True)

Run inference to get the densepose results for a video.

Parameters
video

numpy array (uint8) of a a video in BGR layout with time dimension first or path to a video

VideoLoaderConfig, optional

A video loader config for loading videos (uses all defaults except pix_fmt="bgr24")

bool, optional

Whether to display a progress bar, by default True

Returns

tuple Tuple of (video_array, list of densepose results per frame)

Source code in zamba/models/densepose/densepose_manager.py
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
def predict_video(self, video, video_loader_config=None, pbar=True):
    """Run inference to get the densepose results for a video.

    Parameters
    ----------
    video :
        numpy array (uint8) of a a video in BGR layout with time dimension first or path to a video
    video_loader_config : VideoLoaderConfig, optional
        A video loader config for loading videos (uses all defaults except pix_fmt="bgr24")
    pbar : bool, optional
        Whether to display a progress bar, by default True

    Returns
    -------
    tuple
        Tuple of (video_array, list of densepose results per frame)
    """
    if isinstance(video, (str, Path)):
        video = load_video_frames(video, config=video_loader_config)

    pbar = tqdm if pbar else lambda x, **kwargs: x

    return video, [
        self.predict_image(img)[1] for img in pbar(video, desc="Frames")
    ]  # just the predictions
serialize_image_output(instances, filename = None, write_embeddings = False)

Convert the densepose output into Python-native objects that can be written and read with json.

Parameters
Instance

The output from the densepose model

(str, Path), optional

If not None, the filename to write the output to, by default None

Source code in zamba/models/densepose/densepose_manager.py
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
def serialize_image_output(self, instances, filename=None, write_embeddings=False):
    """Convert the densepose output into Python-native objects that can
        be written and read with json.

    Parameters
    ----------
    instances : Instance
        The output from the densepose model
    filename : (str, Path), optional
        If not None, the filename to write the output to, by default None
    """
    if isinstance(instances, list):
        img_height, img_width = instances[0].image_size
    else:
        img_height, img_width = instances.image_size

    boxes = instances.get("pred_boxes").tensor
    scores = instances.get("scores").tolist()
    labels = instances.get("pred_classes").tolist()

    try:
        pose_result = instances.get("pred_densepose")
    except KeyError:
        pose_result = None

    # include embeddings + segmentation if they exist and they are requested
    write_embeddings = write_embeddings and (pose_result is not None)

    serialized = {
        "instances": [
            {
                "img_height": img_height,
                "img_width": img_width,
                "box": boxes[i].cpu().tolist(),
                "score": scores[i],
                "label": {
                    "value": labels[i],
                    "mesh_name": self.vis_class_to_mesh_name[labels[i]],
                },
                "embedding": pose_result.embedding[[i], ...].cpu().tolist()
                if write_embeddings
                else None,
                "segmentation": pose_result.coarse_segm[[i], ...].cpu().tolist()
                if write_embeddings
                else None,
            }
            for i in range(len(instances))
        ]
    }

    if filename is not None:
        with Path(filename).open("w") as f:
            json.dump(serialized, f, indent=2)

    return serialized
serialize_video_output(instances, filename = None, write_embeddings = False)
Source code in zamba/models/densepose/densepose_manager.py
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
def serialize_video_output(self, instances, filename=None, write_embeddings=False):
    serialized = {
        "frames": [
            self.serialize_image_output(
                frame_instances, filename=None, write_embeddings=write_embeddings
            )
            for frame_instances in instances
        ]
    }

    if filename is not None:
        with Path(filename).open("w") as f:
            json.dump(serialized, f, indent=2)

    return serialized
visualize_image(image_arr, outputs, output_path = None)

Visualize the pose information.

Parameters
numpy array (unit8) BGR

The numpy array representing the image.

outputs

The outputs from running DensePoseManager.predict*

str or Path, optional

If not None, write visualization to this path; by default None

Returns

numpy array (unit8) BGR DensePose outputs visualized on top of the image.

Source code in zamba/models/densepose/densepose_manager.py
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
def visualize_image(self, image_arr, outputs, output_path=None):
    """Visualize the pose information.

    Parameters
    ----------
    image_arr : numpy array (unit8) BGR
        The numpy array representing the image.
    outputs :
        The outputs from running DensePoseManager.predict*
    output_path : str or Path, optional
        If not None, write visualization to this path; by default None

    Returns
    -------
    numpy array (unit8) BGR
        DensePose outputs visualized on top of the image.
    """
    bw_image = cv2.cvtColor(image_arr, cv2.COLOR_BGR2GRAY)
    bw_image = np.tile(bw_image[:, :, np.newaxis], [1, 1, 3])
    data = self.vis_extractor(outputs)
    image_vis = self.visualizer.visualize(bw_image, data)

    if output_path is not None:
        cv2.imwrite(str(output_path), image_vis)

    return image_vis
visualize_video(video_arr, outputs, output_path = None, frame_size = None, fps = 30, pbar = True)

Visualize the pose information on a video

Parameters
numpy array (unit8) BGR, time first

The numpy array representing the video.

outputs

The outputs from running DensePoseManager.predict*

str or Path, optional

If not None, write visualization to this path (should be .mp4); by default None

(innt, float), optional

If frame_size is float, scale up or down by that float value; if frame_size is an integer, set width to that size and scale height appropriately.

int

frames per second for output video if writing; defaults to 30

bool

display a progress bar

Returns

numpy array (unit8) BGR DensePose outputs visualized on top of the image.

Source code in zamba/models/densepose/densepose_manager.py
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
def visualize_video(
    self, video_arr, outputs, output_path=None, frame_size=None, fps=30, pbar=True
):
    """Visualize the pose information on a video

    Parameters
    ----------
    video_arr : numpy array (unit8) BGR, time first
        The numpy array representing the video.
    outputs :
        The outputs from running DensePoseManager.predict*
    output_path : str or Path, optional
        If not None, write visualization to this path (should be .mp4); by default None
    frame_size : (innt, float), optional
        If frame_size is float, scale up or down by that float value; if frame_size is an integer,
        set width to that size and scale height appropriately.
    fps : int
        frames per second for output video if writing; defaults to 30
    pbar : bool
        display a progress bar

    Returns
    -------
    numpy array (unit8) BGR
        DensePose outputs visualized on top of the image.
    """
    pbar = tqdm if pbar else lambda x, **kwargs: x

    out_frames = np.array(
        [
            self.visualize_image(
                image_arr,
                output,
            )
            for image_arr, output in pbar(
                zip(video_arr, outputs), total=video_arr.shape[0], desc="Visualize frames"
            )
        ]
    )

    if output_path is not None:
        # get new size for output video if scaling
        if frame_size is None:
            frame_size = video_arr.shape[2]  # default to same size

        # if float, scale as a multiple
        if isinstance(frame_size, float):
            frame_width = round(video_arr.shape[2] * frame_size)
            frame_height = round(video_arr.shape[1] * frame_size)

        # if int, use as width of the video and scale height proportionally
        elif isinstance(frame_size, int):
            frame_width = frame_size
            scale = frame_width / video_arr.shape[2]
            frame_height = round(video_arr.shape[1] * scale)

        # setup output for writing
        output_path = output_path.with_suffix(".mp4")
        out = cv2.VideoWriter(
            str(output_path),
            cv2.VideoWriter_fourcc(*"mp4v"),
            max(1, int(fps)),
            (frame_width, frame_height),
        )

        for f in pbar(out_frames, desc="Write frames"):
            if (f.shape[0] != frame_height) or (f.shape[1] != frame_width):
                f = cv2.resize(
                    f,
                    (frame_width, frame_height),
                    # https://stackoverflow.com/a/51042104/1692709
                    interpolation=(
                        cv2.INTER_LINEAR if f.shape[1] < frame_width else cv2.INTER_AREA
                    ),
                )
            out.write(f)

        out.release()

    return out_frames

Functions