update

configs/.ipynb_checkpoints/backbone-checkpoint.yaml

@@ -0,0 +1,19 @@
+channels:
+  mobilenetv3_large_100: [16,24,40,112,160]
+  mobilenetv2_120d: [24,32,40,112,192]
+  mobilenetv2_100: [16,24,32,96,160]
+  mnasnet_100: [16,24,40,96,192]
+  efficientnet_b0: [16,24,40,112,192]
+  efficientnet_b3a: [24,32,48,136,232]
+  mixnet_xl: [40,48,64,192,320]
+  dla34: [32,64,128,256,512]
+
+layers:
+  mobilenetv3_large_100: [1,2,3,5,6]
+  mobilenetv2_120d: [1,2,3,5,6]
+  mobilenetv2_100: [1,2,3,5,6]
+  mnasnet_100: [1,2,3,5,6]
+  efficientnet_b0: [1,2,3,5,6]
+  efficientnet_b3a: [1,2,3,5,6]
+  mixnet_xl: [1,2,3,5,6]
+  dla34: [1,2,3,5,6]

dataloaders/stereo/.ipynb_checkpoints/KITTILoader-checkpoint.py

@@ -0,0 +1,176 @@
+import os
+import torch
+import torch.utils.data as data
+import torch
+import torchvision.transforms as transforms
+import random
+from albumentations import Compose, OneOf
+from PIL import Image, ImageOps
+import numpy as np
+from . import preprocess 
+from .stereo_albumentation import RandomShiftRotate, GaussNoiseStereo, RGBShiftStereo, \
+    RandomBrightnessContrastStereo, random_crop, horizontal_flip
+from . import transforms
+from .transforms import RandomColor
+import cv2
+import pdb
+
+IMG_EXTENSIONS = [
+    '.jpg', '.JPG', '.jpeg', '.JPEG',
+    '.png', '.PNG', '.ppm', '.PPM', '.bmp', '.BMP',
+]
+
+def is_image_file(filename):
+    return any(filename.endswith(extension) for extension in IMG_EXTENSIONS)
+
+def default_loader(path):
+    return cv2.imread(path)
+    #return Image.open(path).convert('RGB')
+
+def disparity_loader(path):
+    return Image.open(path)
+
+
+class ImageLoader(data.Dataset):
+    def __init__(self, left, right, left_disparity, calib, th=256, tw=512, shift=0, training=True, loader=default_loader, dploader= disparity_loader):
+
+        self.left = left
+        self.right = right
+        self.disp_L = left_disparity
+        self.calib = calib
+        self.loader = loader
+        self.dploader = dploader
+        self.training = training
+        self.th = th
+        self.tw = tw
+        self.shift = shift
+
+    def __getitem__(self, index):
+        batch = dict()
+
+        left  = self.left[index]
+        right = self.right[index]
+        disp_L= self.disp_L[index]
+        calib = self.calib[index]
+
+        left_img = self.loader(left)
+        right_img = self.loader(right)
+        dataL = self.dploader(disp_L)
+        file = open(calib,"r")
+        cal = file.read()
+        if calib.find('kitti15')==-1:
+            P2 = np.array(cal.split('\n')[2].split(' ')[1:]).astype(np.float32)
+            P3 = np.array(cal.split('\n')[3].split(' ')[1:]).astype(np.float32)
+        else:
+            P2 = np.array(cal.split('\n')[-10].split(' ')[1:]).astype(np.float32)
+            P3 = np.array(cal.split('\n')[-2].split(' ')[1:]).astype(np.float32)
+        P2 = P2.reshape(3,4)
+        P3 = P3.reshape(3,4)
+
+        calib = self.kitti_calib(P2,P3)
+
+        dataL = np.ascontiguousarray(dataL,dtype=np.float32)/256
+
+        # if 'kitti15' in left:
+        #     disp_R = disp_L.replace('occ_0','occ_1')
+        #     dataR = self.dploader(disp_R)
+        #     dataR = np.ascontiguousarray(dataR,dtype=np.float32)/256
+
+        if self.training:  
+            # if 'kitti15' in left:
+            #     left_img, right_img, dataL = horizontal_flip(left_img, right_img, dataL, dataR)
+
+            pad_h, pad_w = 384-left_img.shape[0], 1280-left_img.shape[1]
+
+            left_img = np.pad(left_img,((0,pad_h),(0,pad_w),(0,0)))
+            right_img = np.pad(right_img,((0,pad_h),(0,pad_w),(0,0)))
+
+            h,w,_ = left_img.shape
+            th, tw = self.th, self.tw
+
+            shift = random.randint(-self.shift,self.shift)
+            x1 = random.randint(0, w - tw)
+            y1 = random.randint(0, h - th)
+
+            # if x1 + shift < 0 or  x1 + shift + tw > w:
+            shift = 0
+
+            left_img_raw = left_img[y1:y1+th,x1+shift:x1+shift+tw,:]
+            right_img_raw = right_img[y1:y1+th,x1:x1+tw,:]
+
+            imL_lab = cv2.cvtColor(
+                left_img_raw,#cv2.resize(left_img,None,None,0.25,0.25),
+                cv2.COLOR_BGR2LAB)
+
+            dataL = np.pad(dataL[:,:,np.newaxis],((0,pad_h),(0,pad_w),(0,0)))[:,:,0]
+            dataL = dataL[y1:y1 + th, x1 + shift:x1 + tw + shift]
+            dataL = dataL - shift
+
+            img = {'left':left_img_raw,'right':right_img_raw}
+            # img = self.train_aug(img)
+
+            left_img_raw, right_img_raw = img['left'], img['right']
+
+            processed = preprocess.get_transform(augment=False)  
+            left_img   = processed(left_img_raw)
+            right_img  = processed(right_img_raw)
+
+            left_img_raw = np.transpose(left_img_raw,(2,0,1)).astype(np.float32)
+            right_img_raw = np.transpose(right_img_raw,(2,0,1)).astype(np.float32)
+
+            batch['imgL'], batch['imgR'], batch['disp_true'] = left_img, right_img, dataL
+            batch['imgLRaw'], batch['imgRRaw'], batch['imgLLab'] = left_img_raw, right_img_raw, imL_lab
+            batch['calib'], batch['x1'], batch['y1'] = calib, x1, y1
+
+            return batch
+        else:
+            h,w,_ = left_img.shape
+            imL = left_img
+            pad_h, pad_w = 384-h, 1280-w
+
+            # left_img_raw = left_img[h-352:h,w-1216:w,:]
+            # right_img_raw = right_img[h-352:h,w-1216:w,:]
+            left_img_raw = left_img  # np.pad(left_img,((0,pad_h),(0,pad_w),(0,0)))
+            right_img_raw = right_img  # np.pad(right_img,((0,pad_h),(0,pad_w),(0,0)))
+
+            imL_lab = cv2.cvtColor(
+                left_img_raw,#cv2.resize(left_img,None,None,0.25,0.25),
+                cv2.COLOR_BGR2LAB)
+
+            # dataL = dataL.crop((w-1216, h-352, w, h))
+            # dataL = np.pad(dataL,((0,pad_h),(0,pad_w)))
+
+            processed = preprocess.get_transform(augment=False)  
+            left_img       = processed(left_img_raw)
+            right_img      = processed(right_img_raw)
+
+            batch['imgL'], batch['imgR'], batch['disp_true'] = left_img, right_img, dataL
+            batch['imgLLab'] = imL_lab
+            batch['calib'] = calib
+
+            return batch
+
+    def __len__(self):
+        return len(self.left)
+
+    def train_aug(self, img):
+        transformation = Compose([
+                RGBShiftStereo(always_apply=True, p_asym=0.5),
+                RandomBrightnessContrastStereo(always_apply=True, p_asym=0.5)
+                ])
+        return transformation(**img)
+
+        # transformation = transforms.Compose([
+        #         RandomColor()
+        #         ])
+        # return transformation(img)
+
+    def kitti_calib(self, P2, P3):
+        t2 = np.array([P2[0,-1]/P2[0,0],P2[1,-1]/P2[1,1],P2[2,-1]])
+        t3 = np.array([P3[0,-1]/P3[0,0],P3[1,-1]/P3[1,1],P3[2,-1]])
+        t = t2-t3
+        baseline = np.linalg.norm(t,2)
+
+        K = P2[:,:-1]
+
+        return {'K':K,'baseline':baseline}