feature_keynet.py

"""
* This file is part of PYSLAM 
*
* Adpated from https://raw.githubusercontent.com/axelBarroso/Key.Net/master/extract_multiscale_features.py, see the license therein. 
* Copyright (C) 2016-present Luigi Freda <luigi dot freda at gmail dot com> 
*
* PYSLAM is free software: you can redistribute it and/or modify
* it under the terms of the GNU General Public License as published by
* the Free Software Foundation, either version 3 of the License, or
* (at your option) any later version.
*
* PYSLAM is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License for more details.
*
* You should have received a copy of the GNU General Public License
* along with PYSLAM. If not, see <http://www.gnu.org/licenses/>.
"""

import config
config.cfg.set_lib('keynet') 

import warnings # to disable tensorflow-numpy warnings: from https://github.com/tensorflow/tensorflow/issues/30427
warnings.filterwarnings('ignore', category=FutureWarning)

import os, sys, cv2
#sys.path.append(os.path.dirname(os.path.dirname(os.path.realpath(__file__))))
from os import path, mkdir
import argparse
import keyNet.aux.tools as aux
from skimage.transform import pyramid_gaussian
import HSequences_bench.tools.geometry_tools as geo_tools
import HSequences_bench.tools.repeatability_tools as rep_tools
from keyNet.model.keynet_architecture import *
import keyNet.aux.desc_aux_function as loss_desc
from keyNet.model.hardnet_pytorch import *
from keyNet.datasets.dataset_utils import read_bw_image
import torch

from threading import RLock

from utils_tf import set_tf_logging
from utils_sys import Printer, print_options


kVerbose = True   


def build_keynet_config(keynet_base_path):
    parser = argparse.ArgumentParser(description='HSequences Extract Features')

    # parser.add_argument('--list-images', type=str, help='File containing the image paths for extracting features.',
    #                     required=True)

    # parser.add_argument('--results-dir', type=str, default='extracted_features/',
    #                     help='The output path to save the extracted keypoint.')

    parser.add_argument('--network-version', type=str, default='KeyNet_default',
                        help='The Key.Net network version name')

    parser.add_argument('--checkpoint-det-dir', type=str, default=keynet_base_path + 'keyNet/pretrained_nets/KeyNet_default',
                        help='The path to the checkpoint file to load the detector weights.')

    parser.add_argument('--pytorch-hardnet-dir', type=str, default=keynet_base_path + 'keyNet/pretrained_nets/HardNet++.pth',
                        help='The path to the checkpoint file to load the HardNet descriptor weights.')

    # Detector Settings

    parser.add_argument('--num-filters', type=int, default=8,
                        help='The number of filters in each learnable block.')

    parser.add_argument('--num-learnable-blocks', type=int, default=3,
                        help='The number of learnable blocks after handcrafted block.')

    parser.add_argument('--num-levels-within-net', type=int, default=3,
                        help='The number of pyramid levels inside the architecture.')

    parser.add_argument('--factor-scaling-pyramid', type=float, default=1.2,
                        help='The scale factor between the multi-scale pyramid levels in the architecture.')

    parser.add_argument('--conv-kernel-size', type=int, default=5,
                        help='The size of the convolutional filters in each of the learnable blocks.')

    # Multi-Scale Extractor Settings

    parser.add_argument('--extract-MS', type=bool, default=True,
                        help='Set to True if you want to extract multi-scale features.')

    parser.add_argument('--num-points', type=int, default=2000,
                        help='The number of desired features to extract.')

    parser.add_argument('--nms-size', type=int, default=15,
                        help='The NMS size for computing the validation repeatability.')

    parser.add_argument('--border-size', type=int, default=15,
                        help='The number of pixels to remove from the borders to compute the repeatability.')

    parser.add_argument('--order-coord', type=str, default='xysr',
                        help='The coordinate order that follows the extracted points. Use yxsr or xysr.')

    parser.add_argument('--random-seed', type=int, default=12345,
                        help='The random seed value for TensorFlow and Numpy.')

    parser.add_argument('--pyramid_levels', type=int, default=5,
                        help='The number of downsample levels in the pyramid.')

    parser.add_argument('--upsampled-levels', type=int, default=1,
                        help='The number of upsample levels in the pyramid.')

    parser.add_argument('--scale-factor-levels', type=float, default=np.sqrt(2),
                        help='The scale factor between the pyramid levels.')

    parser.add_argument('--scale-factor', type=float, default=2.,
                        help='The scale factor to extract patches before descriptor.')

    # GPU Settings

    parser.add_argument('--gpu-memory-fraction', type=float, default=0.3,
                        help='The fraction of GPU used by the script.')

    parser.add_argument('--gpu-visible-devices', type=str, default="0",
                        help='Set CUDA_VISIBLE_DEVICES variable.')

    args = parser.parse_known_args()[0]

    # remove verbose bits from tf
    # os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
    # tf.logging.set_verbosity(tf.logging.ERROR)

    # Set CUDA GPU environment
    os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu_visible_devices

    print('Using KeyNet version:' + args.network_version)

    if not args.extract_MS:
        args.pyramid_levels = 0
        args.upsampled_levels = 0
        
    return args 
                            
                
# convert matrix of pts into list of keypoints
def convert_pts_to_keypoints(pts, scores, sizes, levels): 
    assert(len(pts)==len(scores))
    kps = []
    if pts is not None: 
        # convert matrix [Nx2] of pts into list of keypoints  
        kps = [ cv2.KeyPoint(p[0], p[1], _size=sizes[i], _response=scores[i], _octave=levels[i]) for i,p in enumerate(pts) ]                      
    return kps         


# interface for pySLAM 
class KeyNetDescFeature2D: 
    def __init__(self,       
                 num_features=2000,
                 num_levels=5,                   # The number of downsample levels in the pyramid.
                 scale_factor=2,                 # The scale factor to extract patches before descriptor.
                 scale_factor_levels=np.sqrt(2), # The scale factor between the pyramid levels.
                 do_cuda=True,
                 do_tf_logging=False):  
        print('Using KeyNetDescFeature2D')   
        self.lock = RLock()
        self.model_base_path = config.cfg.root_folder + '/thirdparty/keynet/'
        
        set_tf_logging(do_tf_logging)
        
        self.do_cuda = do_cuda & torch.cuda.is_available()
        print('cuda:',self.do_cuda)     
        device = torch.device("cuda:0" if self.do_cuda else "cpu")   
        
        self.session = None 
        
        self.keypoint_size = 8  # just a representative size for visualization and in order to convert extracted points to cv2.KeyPoint        

        self.pts = []
        self.kps = []        
        self.des = []
        self.scales = []
        self.scores = []        
        self.frame = None 
        
        keynet_config = build_keynet_config(self.model_base_path)
        self.keynet_config = keynet_config       
        keynet_config.num_points = num_features
        keynet_config.pyramid_levels = num_levels
        keynet_config.scale_factor = scale_factor
        keynet_config.scale_factor_levels = scale_factor_levels
        
        print_options(self.keynet_config,'KEYNET CONFIG')
        
        print('==> Loading pre-trained network.') 
        self.load_model()      
        print('==> Successfully loaded pre-trained network.')

    @property
    def num_features(self):
        return self.keynet_config.num_points
    
    @property
    def num_levels(self):
        return self.keynet_config.pyramid_levels    
    
    @property
    def scale_factor(self):
        return self.keynet_config.scale_factor


    def __del__(self): 
        self.close()
        
        
    def close(self):
        if self.session is not None:
            print('KEYNET: closing tf session')
            self.session.close()
            tf.reset_default_graph()    
               
    
    def load_model(self):        
        # Create graph before session :)
        self.graph = tf.Graph().as_default()
        
        # GPU Usage
        tf_config = tf.ConfigProto()
        tf_config.gpu_options.per_process_gpu_memory_fraction = self.keynet_config.gpu_memory_fraction
        tf_config.gpu_options.allow_growth = True

        #with tf.Session(config=config) as sess:        
        self.session = tf.Session(config=tf_config)         

        tf.set_random_seed(self.keynet_config.random_seed)

        with tf.name_scope('inputs'):

            # Define the input tensor shape
            tensor_input_shape = (None, None, None, 1)

            self.input_network = tf.placeholder(dtype=tf.float32, shape=tensor_input_shape, name='input_network')
            self.dimension_image = tf.placeholder(dtype=tf.int32, shape=(3,), name='dimension_image')
            self.kpts_coord = tf.placeholder(dtype=tf.float32, shape=(None, 2), name='kpts_coord')
            self.kpts_batch = tf.placeholder(dtype=tf.int32, shape=(None,), name='kpts_batch')
            self.kpts_scale = tf.placeholder(dtype=tf.float32, name='kpts_scale')
            self.phase_train = tf.placeholder(tf.bool, name='phase_train')

        with tf.name_scope('model_deep_detector'):

            deep_architecture = keynet(self.keynet_config)
            output_network = deep_architecture.model(self.input_network, self.phase_train, self.dimension_image, reuse=False)
            self.maps = tf.nn.relu(output_network['output'])

        # Extract Patches from inputs:
        self.input_patches = loss_desc.build_patch_extraction(self.kpts_coord, self.kpts_batch, self.input_network, kpts_scale=self.kpts_scale)

        # Define Pytorch HardNet
        self.model = HardNet()
        checkpoint = torch.load(self.keynet_config.pytorch_hardnet_dir)
        self.model.load_state_dict(checkpoint['state_dict'])
        if self.do_cuda:
            self.model.cuda()
            print('Extracting torch model on GPU')
        else:
            print('Extracting torch model on CPU')
            self.model = model.cpu()        
        self.model.eval()         

        # Define variables
        detect_var = [v for v in tf.trainable_variables(scope='model_deep_detector')]

        if os.listdir(self.keynet_config.checkpoint_det_dir):
            init_assign_op_det, init_feed_dict_det = tf_contrib.framework.assign_from_checkpoint(
                tf.train.latest_checkpoint(self.keynet_config.checkpoint_det_dir), detect_var)

        point_level = []
        tmp = 0.0
        factor_points = (self.keynet_config.scale_factor_levels ** 2)
        self.levels = self.keynet_config.pyramid_levels + self.keynet_config.upsampled_levels + 1
        #print('levels: ', [i for i in range(self.levels)])
        for idx_level in range(self.levels):
            tmp += factor_points ** (-1 * (idx_level - self.keynet_config.upsampled_levels))
            point_level.append(self.keynet_config.num_points * factor_points ** (-1 * (idx_level - self.keynet_config.upsampled_levels)))

        self.point_level = np.asarray(list(map(lambda x: int(x / tmp), point_level)))
        #print('self.point_level:',self.point_level)

        self.session.run(tf.global_variables_initializer())

        if os.listdir(self.keynet_config.checkpoint_det_dir):
            self.session.run(init_assign_op_det, init_feed_dict_det)        
        
            
    def extract_keynet_features(self, image):
        pyramid = pyramid_gaussian(image, max_layer=self.keynet_config.pyramid_levels, downscale=self.keynet_config.scale_factor_levels)

        score_maps = {}
        for (j, resized) in enumerate(pyramid):
            im = resized.reshape(1, resized.shape[0], resized.shape[1], 1)

            feed_dict = {
                self.input_network: im,
                self.phase_train: False,
                self.dimension_image: np.array([1, im.shape[1], im.shape[2]], dtype=np.int32),
            }

            im_scores = self.session.run(self.maps, feed_dict=feed_dict)

            im_scores = geo_tools.remove_borders(im_scores, borders=self.keynet_config.border_size)
            score_maps['map_' + str(j + 1 + self.keynet_config.upsampled_levels)] = im_scores[0, :, :, 0]

        if self.keynet_config.upsampled_levels:
            for j in range(self.keynet_config.upsampled_levels):
                factor = self.keynet_config.scale_factor_levels ** (self.keynet_config.upsampled_levels - j)
                up_image = cv2.resize(image, (0, 0), fx=factor, fy=factor)

                im = np.reshape(up_image, (1, up_image.shape[0], up_image.shape[1], 1))

                feed_dict = {
                    self.input_network: im,
                    self.phase_train: False,
                    self.dimension_image: np.array([1, im.shape[1], im.shape[2]], dtype=np.int32),
                }

                im_scores = self.session.run(self.maps, feed_dict=feed_dict)

                im_scores = geo_tools.remove_borders(im_scores, borders=self.keynet_config.border_size)
                score_maps['map_' + str(j + 1)] = im_scores[0, :, :, 0]

        im_pts = []
        im_pts_levels = []        
        for idx_level in range(self.levels):

            scale_value = (self.keynet_config.scale_factor_levels ** (idx_level - self.keynet_config.upsampled_levels))
            scale_factor = 1. / scale_value

            h_scale = np.asarray([[scale_factor, 0., 0.], [0., scale_factor, 0.], [0., 0., 1.]])
            h_scale_inv = np.linalg.inv(h_scale)
            h_scale_inv = h_scale_inv / h_scale_inv[2, 2]

            num_points_level = self.point_level[idx_level]
            #print('num_points_level:',num_points_level)
            if idx_level > 0:
                res_points = int(np.asarray([self.point_level[a] for a in range(0, idx_level + 1)]).sum() - len(im_pts))
                num_points_level = res_points

            im_scores = rep_tools.apply_nms(score_maps['map_' + str(idx_level + 1)], self.keynet_config.nms_size)
            im_pts_tmp = geo_tools.get_point_coordinates(im_scores, num_points=num_points_level, order_coord='xysr')

            im_pts_tmp = geo_tools.apply_homography_to_points(im_pts_tmp, h_scale_inv)

            if not idx_level:
                im_pts = im_pts_tmp
            else:
                im_pts = np.concatenate((im_pts, im_pts_tmp), axis=0)
                
            im_pts_levels_tmp = np.ones(len(im_pts),dtype=np.int32)*idx_level     
            im_pts_levels = np.concatenate((im_pts_levels, im_pts_levels_tmp), axis=0).astype(np.int32)                             

        if self.keynet_config.order_coord == 'yxsr':
            im_pts = np.asarray(list(map(lambda x: [x[1], x[0], x[2], x[3]], im_pts)))

        sorted_idxs = (-1 * im_pts[:, 3]).argsort() # sort points with their scores 
        im_pts = im_pts[sorted_idxs]
        im_pts_levels = im_pts_levels[sorted_idxs]     
        #print('im_pts_levels:',im_pts_levels)   
        im_pts = im_pts[:self.keynet_config.num_points]        
        im_pts_levels = im_pts_levels[:self.keynet_config.num_points]

        # Extract descriptor from features
        descriptors = []
        im = image.reshape(1, image.shape[0], image.shape[1], 1)
        for idx_desc_batch in range(int(len(im_pts) / 250 + 1)):
            points_batch = im_pts[idx_desc_batch * 250: (idx_desc_batch + 1) * 250]

            if not len(points_batch):
                break

            feed_dict = {
                self.input_network: im,
                self.phase_train: False,
                self.kpts_coord: points_batch[:, :2],
                self.kpts_scale: self.keynet_config.scale_factor * points_batch[:, 2],
                self.kpts_batch: np.zeros(len(points_batch)),
                self.dimension_image: np.array([1, im.shape[1], im.shape[2]], dtype=np.int32),
            }

            patch_batch = self.session.run(self.input_patches, feed_dict=feed_dict)
            patch_batch = np.reshape(patch_batch, (patch_batch.shape[0], 1, 32, 32))
            data_a = torch.from_numpy(patch_batch)
            data_a = data_a.cuda()
            data_a = Variable(data_a)
            with torch.no_grad():
                out_a = self.model(data_a)
            desc_batch = out_a.data.cpu().numpy().reshape(-1, 128)
            if idx_desc_batch == 0:
                descriptors = desc_batch
            else:
                descriptors = np.concatenate([descriptors, desc_batch], axis=0)

        return im_pts, descriptors, im_pts_levels           
        
        
    def compute_kps_des(self, im):
        with self.lock:         

            im = im.astype(float) / im.max()
            im_pts, descriptors, im_pts_levels = self.extract_keynet_features(im)
                        
            self.pts = im_pts[:,:2]    
            scales = im_pts[:,2]        
            scores = im_pts[:,3]      
            pts_levels = im_pts_levels           
            
            #print('scales:',self.scales)
            
            self.kps = convert_pts_to_keypoints(self.pts, scores, scales*self.keypoint_size, pts_levels)
            
            return self.kps, descriptors   
        
           
    def detectAndCompute(self, frame, mask=None):  #mask is a fake input  
        with self.lock: 
            self.frame = frame         
            self.kps, self.des = self.compute_kps_des(frame)            
            if kVerbose:
                print('detector: KEYNET, descriptor: KEYNET, #features: ', len(self.kps), ', frame res: ', frame.shape[0:2])                  
            return self.kps, self.des
    
           
    # return keypoints if available otherwise call detectAndCompute()    
    def detect(self, frame, mask=None):  # mask is a fake input  
        with self.lock:         
            #if self.frame is not frame:
            self.detectAndCompute(frame)        
            return self.kps
    
    
    # return descriptors if available otherwise call detectAndCompute()  
    def compute(self, frame, kps=None, mask=None): # kps is a fake input, mask is a fake input
        with self.lock:         
            if self.frame is not frame:
                Printer.orange('WARNING: KEYNET is recomputing both kps and des on last input frame', frame.shape)            
                self.detectAndCompute(frame)
            return self.kps, self.des   
    
    
    # return descriptors if available otherwise call detectAndCompute()  
    def compute(self, frame, kps=None, mask=None): # kps is a fake input, mask is a fake input
        with self.lock:         
            if self.frame is not frame:
                #Printer.orange('WARNING: KEYNET is recomputing both kps and des on last input frame', frame.shape)            
                self.detectAndCompute(frame)
            return self.kps, self.des