Source code for wbia.algo.detect.fasterrcnn

# -*- coding: utf-8 -*-
"""
Interface to Faster R-CNN object proposals.
"""
import logging
import utool as ut
import vtool as vt
from os.path import abspath, dirname, expanduser, join, exists  # NOQA
import numpy as np
import sys
import cv2

(print, rrr, profile) = ut.inject2(__name__, '[faster r-cnn]')
logger = logging.getLogger('wbia')

# SCRIPT_PATH = abspath(dirname(__file__))
SCRIPT_PATH = abspath(expanduser(join('~', 'code', 'py-faster-rcnn')))

if not ut.get_argflag('--no-faster-rcnn'):
    try:
        assert exists(SCRIPT_PATH)

        def add_path(path):
            # if path not in sys.path:
            sys.path.insert(0, path)

        # Add pycaffe to PYTHONPATH
        pycaffe_path = join(SCRIPT_PATH, 'caffe-fast-rcnn', 'python')
        add_path(pycaffe_path)

        # Add caffe lib path to PYTHONPATH
        lib_path = join(SCRIPT_PATH, 'lib')
        add_path(lib_path)

        import caffe

        ut.reload_module(caffe)
        from fast_rcnn.config import cfg
        from fast_rcnn.test import im_detect

        # from fast_rcnn.nms_wrapper import nms
    except AssertionError:
        logger.info(
            'WARNING Failed to find py-faster-rcnn. ' 'Faster R-CNN is unavailable'
        )
        # if ut.SUPER_STRICT:
        #     raise
    except ImportError:
        logger.info('WARNING Failed to import fast_rcnn. ' 'Faster R-CNN is unavailable')
        # if ut.SUPER_STRICT:
        #     raise


VERBOSE_SS = ut.get_argflag('--verbdss') or ut.VERBOSE


CONFIG_URL_DICT = {
    # 'pretrained-fast-vgg-pascal' : 'https://wildbookiarepository.azureedge.net/models/pretrained.fastrcnn.vgg16.pascal.prototxt',  # Trained on PASCAL VOC 2007
    'pretrained-vgg-pascal': 'https://wildbookiarepository.azureedge.net/models/pretrained.fasterrcnn.vgg16.pascal.prototxt',  # Trained on PASCAL VOC 2007
    'pretrained-zf-pascal': 'https://wildbookiarepository.azureedge.net/models/pretrained.fasterrcnn.zf.pascal.prototxt',  # Trained on PASCAL VOC 2007
    'pretrained-vgg-ilsvrc': 'https://wildbookiarepository.azureedge.net/models/pretrained.fasterrcnn.vgg16.ilsvrc.prototxt',  # Trained on ILSVRC 2014
    'pretrained-zf-ilsvrc': 'https://wildbookiarepository.azureedge.net/models/pretrained.fasterrcnn.zf.ilsvrc.prototxt',  # Trained on ILSVRC 2014
    'default': 'https://wildbookiarepository.azureedge.net/models/pretrained.fasterrcnn.vgg16.pascal.prototxt',  # Trained on PASCAL VOC 2007
    None: 'https://wildbookiarepository.azureedge.net/models/pretrained.fasterrcnn.vgg16.pascal.prototxt',  # Trained on PASCAL VOC 2007
}


def _parse_weight_from_cfg(url):
    return url.replace('.prototxt', '.caffemodel')


def _parse_classes_from_cfg(url):
    return url.replace('.prototxt', '.classes')


def _parse_class_list(classes_filepath):
    # Load classes from file into the class list
    assert exists(classes_filepath)
    class_list = []
    with open(classes_filepath) as classes:
        for line in classes.readlines():
            line = line.strip()
            if len(line) > 0:
                class_list.append(line)
    return class_list


[docs]def detect_gid_list(ibs, gid_list, downsample=True, verbose=VERBOSE_SS, **kwargs): """ Args: gid_list (list of int): the list of IBEIS image_rowids that need detection downsample (bool, optional): a flag to indicate if the original image sizes should be used; defaults to True True: ibs.get_image_detectpaths() is used False: ibs.get_image_paths() is used Kwargs (optional): refer to the Faster R-CNN documentation for configuration settings Args: ibs (wbia.IBEISController): image analysis api gid_list (list of int): the list of IBEIS image_rowids that need detection downsample (bool, optional): a flag to indicate if the original image sizes should be used; defaults to True Kwargs: detector, config_filepath, weights_filepath, verbose Yields: tuple: (gid, gpath, result_list) CommandLine: python -m wbia.algo.detect.fasterrcnn detect_gid_list --show Example: >>> # DISABLE_DOCTEST >>> from wbia.algo.detect.fasterrcnn import * # NOQA >>> from wbia.core_images import LocalizerConfig >>> import wbia >>> ibs = wbia.opendb('testdb1') >>> gid_list = ibs.get_valid_gids() >>> config = {'verbose': True} >>> downsample = False >>> results_list = detect_gid_list(ibs, gid_list, downsample, **config) >>> results_list = list(results_list) >>> print('result lens = %r' % (map(len, list(results_list)))) >>> print('result[0] = %r' % (len(list(results_list[0][2])))) >>> config = {'verbose': True} >>> downsample = False >>> results_list = detect_gid_list(ibs, gid_list, downsample, **config) >>> results_list = list(results_list) >>> print('result lens = %r' % (map(len, list(results_list)))) >>> print('result[0] = %r' % (len(list(results_list[0][2])))) >>> ut.quit_if_noshow() >>> import wbia.plottool as pt >>> ut.show_if_requested() Yields: results (list of dict) """ # Get new gpaths if downsampling if downsample: gpath_list = ibs.get_image_detectpaths(gid_list) neww_list = [vt.open_image_size(gpath)[0] for gpath in gpath_list] oldw_list = [oldw for (oldw, oldh) in ibs.get_image_sizes(gid_list)] downsample_list = [oldw / neww for oldw, neww in zip(oldw_list, neww_list)] orient_list = [1] * len(gid_list) else: gpath_list = ibs.get_image_paths(gid_list) downsample_list = [None] * len(gpath_list) orient_list = ibs.get_image_orientation(gid_list) # Run detection results_iter = detect(gpath_list, verbose=verbose, **kwargs) # Upscale the results _iter = zip(downsample_list, gid_list, orient_list, results_iter) for downsample, gid, orient, (gpath, result_list) in _iter: # Upscale the results back up to the original image size for result in result_list: if downsample is not None and downsample != 1.0: for key in ['xtl', 'ytl', 'width', 'height']: result[key] = int(result[key] * downsample) bbox = ( result['xtl'], result['ytl'], result['width'], result['height'], ) bbox_list = [bbox] bbox = bbox_list[0] result['xtl'], result['ytl'], result['width'], result['height'] = bbox yield (gid, gpath, result_list)
[docs]def detect( gpath_list, config_filepath, weight_filepath, class_filepath, sensitivity, verbose=VERBOSE_SS, use_gpu=True, use_gpu_id=0, **kwargs, ): """ Args: gpath_list (list of str): the list of image paths that need proposal candidates Kwargs (optional): refer to the Faster R-CNN documentation for configuration settings Returns: iter """ cfg.TEST.HAS_RPN = True # Use RPN for proposals # Get correct config if specified with shorthand config_url = None if config_filepath in CONFIG_URL_DICT: config_url = CONFIG_URL_DICT[config_filepath] config_filepath = ut.grab_file_url(config_url, appname='wbia', check_hash=True) # Get correct weights if specified with shorthand if weight_filepath in CONFIG_URL_DICT: if weight_filepath is None and config_url is not None: config_url_ = config_url else: config_url_ = CONFIG_URL_DICT[weight_filepath] weight_url = _parse_weight_from_cfg(config_url_) weight_filepath = ut.grab_file_url(weight_url, appname='wbia', check_hash=True) if class_filepath is None: class_url = _parse_classes_from_cfg(config_url) class_filepath = ut.grab_file_url( class_url, appname='wbia', check_hash=True, verbose=verbose ) class_list = _parse_class_list(class_filepath) # Need to convert unicode strings to Python strings to support Boost Python # call signatures in caffe prototxt_filepath = str(config_filepath) # alias to Caffe nomenclature caffemodel_filepath = str(weight_filepath) # alias to Caffe nomenclature assert exists(prototxt_filepath), 'Specified prototxt file not found' assert exists(caffemodel_filepath), 'Specified caffemodel file not found' if use_gpu: caffe.set_mode_gpu() caffe.set_device(use_gpu_id) cfg.GPU_ID = use_gpu_id else: caffe.set_mode_cpu() net = caffe.Net(prototxt_filepath, caffemodel_filepath, caffe.TEST) # Warm-up network on a dummy image im = 128 * np.ones((300, 500, 3), dtype=np.uint8) for i in range(2): _, _ = im_detect(net, im) results_list_ = [] for gpath in gpath_list: image = cv2.imread(gpath) score_list, bbox_list = im_detect(net, image) # Compile results result_list_ = [] for class_index, class_name in enumerate(class_list[1:]): class_index += 1 # because we skipped background class_boxes = bbox_list[:, 4 * class_index : 4 * (class_index + 1)] class_scores = score_list[:, class_index] dets_list = np.hstack((class_boxes, class_scores[:, np.newaxis])) dets_list = dets_list.astype(np.float32) # # Perform NMS # keep_list = nms(dets_list, nms_sensitivity) # dets_list = dets_list[keep_list, :] # Perform sensitivity check keep_list = np.where(dets_list[:, -1] >= sensitivity)[0] dets_list = dets_list[keep_list, :] for (xtl, ytl, xbr, ybr, conf) in dets_list: xtl = int(np.around(xtl)) ytl = int(np.around(ytl)) xbr = int(np.around(xbr)) ybr = int(np.around(ybr)) confidence = float(conf) result_dict = { 'xtl': xtl, 'ytl': ytl, 'width': xbr - xtl, 'height': ybr - ytl, 'class': class_name, 'confidence': confidence, } result_list_.append(result_dict) results_list_.append(result_list_) results_list = zip(gpath_list, results_list_) return results_list