Fixed stuff

natanielruiz

2017-10-30 af51d0ecb51ad4d6c8ed086855bd3c411ebc4aa0

 code/test_on_video.py

@@ -25,6 +25,7 @@
    parser.add_argument('--video', dest='video_path', help='Path of video')
    parser.add_argument('--bboxes', dest='bboxes', help='Bounding box annotations of frames')
    parser.add_argument('--output_string', dest='output_string', help='String appended to output file')
    parser.add_argument('--n_frames', dest='n_frames', help='Number of frames', type=int)
    args = parser.parse_args()
    return args

@@ -48,9 +49,7 @@
    # ResNet101 with 3 outputs.
    # model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 23, 3], 66)
    # ResNet50
    model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 66)
    # ResNet18
    # model = hopenet.Hopenet(torchvision.models.resnet.BasicBlock, [2, 2, 2, 2], 66)
    model = hopenet.Hopenet(torchvision.models.resnet.Bottleneck, [3, 4, 6, 3], 66, 0)

    print 'Loading snapshot.'
    # Load snapshot
@@ -60,7 +59,8 @@
    print 'Loading data.'

    transformations = transforms.Compose([transforms.Scale(224),
    transforms.RandomCrop(224), transforms.ToTensor()])
    transforms.CenterCrop(224), transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])])

    model.cuda(gpu)

@@ -74,6 +74,8 @@
    idx_tensor = torch.FloatTensor(idx_tensor).cuda(gpu)

    video = cv2.VideoCapture(video_path)

    # New cv2
    width = int(video.get(cv2.CAP_PROP_FRAME_WIDTH))   # float
    height = int(video.get(cv2.CAP_PROP_FRAME_HEIGHT)) # float

@@ -81,13 +83,24 @@
    fourcc = cv2.VideoWriter_fourcc(*'MJPG')
    out = cv2.VideoWriter('output/video/output-%s.avi' % args.output_string, fourcc, 30.0, (width, height))

    # # Old cv2
    # width = int(video.get(cv2.cv.CV_CAP_PROP_FRAME_WIDTH))   # float
    # height = int(video.get(cv2.cv.CV_CAP_PROP_FRAME_HEIGHT)) # float
    #
    # # Define the codec and create VideoWriter object
    # fourcc = cv2.cv.CV_FOURCC(*'MJPG')
    # out = cv2.VideoWriter('output/video/output-%s.avi' % args.output_string, fourcc, 30.0, (width, height))

    txt_out = open('output/video/output-%s.txt' % args.output_string, 'w')

    bbox_file = open(args.bboxes, 'r')
    frame_num = 1

    # TODO: support for several bounding boxes
    for line in bbox_file:
    with open(args.bboxes, 'r') as f:
        bbox_line_list = f.read().splitlines()

    idx = 0
    while idx < len(bbox_line_list):
        line = bbox_line_list[idx]
        line = line.strip('\n')
        line = line.split(' ')
        det_frame_num = int(line[0])
@@ -95,12 +108,8 @@
        print frame_num

        # Stop at a certain frame number
        if frame_num > 10000:
            out.release()
            video.release()
            bbox_file.close()
            txt_out.close()
            sys.exit(0)
        if frame_num > args.n_frames:
            break

        # Save all frames as they are if they don't have bbox annotation.
        while frame_num < det_frame_num:
@@ -108,64 +117,73 @@
            if ret == False:
                out.release()
                video.release()
                bbox_file.close()
                txt_out.close()
                sys.exit(0)
            out.write(frame)
            frame_num += 1

        # Start processing frame with bounding box
        ret,frame = video.read()
        if ret == False:
            out.release()
            video.release()
            bbox_file.close()
            txt_out.close()
            sys.exit(0)
            break

        x_min, y_min, x_max, y_max = int(line[1]), int(line[2]), int(line[3]), int(line[4])
        x_min -= 100
        x_max += 100
        y_min -= 200
        y_max += 50
        x_min = max(x_min, 0)
        y_min = max(y_min, 0)
        x_max = min(frame.shape[1], x_max)
        y_max = min(frame.shape[0], y_max)
        # Crop image
        img = frame[y_min:y_max,x_min:x_max]
        img = Image.fromarray(img)
        while True:
            x_min, y_min, x_max, y_max, conf = int(float(line[1])), int(float(line[2])), int(float(line[3])), int(float(line[4])), float(line[5])

        # Transform
        img = transformations(img)
        img_shape = img.size()
        img = img.view(1, img_shape[0], img_shape[1], img_shape[2])
        img = Variable(img).cuda(gpu)
        yaw, pitch, roll = model(img)
            if conf > 0.98:
                bbox_width = abs(x_max - x_min)
                bbox_height = abs(y_max - y_min)
                # x_min -= 3 * bbox_width / 4
                # x_max += 3 * bbox_width / 4
                # y_min -= 3 * bbox_height / 4
                # y_max += bbox_height / 4
                x_min -= 50
                x_max += 50
                y_min -= 50
                y_max += 30
                x_min = max(x_min, 0)
                y_min = max(y_min, 0)
                x_max = min(frame.shape[1], x_max)
                y_max = min(frame.shape[0], y_max)
                # Crop image
                img = frame[y_min:y_max,x_min:x_max]
                img = Image.fromarray(img)

        yaw_predicted = F.softmax(yaw)
        pitch_predicted = F.softmax(pitch)
        roll_predicted = F.softmax(roll)
        # Get continuous predictions in degrees.
        yaw_predicted = torch.sum(yaw_predicted.data[0] * idx_tensor) * 3 - 99
        pitch_predicted = torch.sum(pitch_predicted.data[0] * idx_tensor) * 3 - 99
        roll_predicted = torch.sum(roll_predicted.data[0] * idx_tensor) * 3 - 99
                # Transform
                img = transformations(img)
                img_shape = img.size()
                img = img.view(1, img_shape[0], img_shape[1], img_shape[2])
                img = Variable(img).cuda(gpu)
                yaw, pitch, roll, angles = model(img)

        # Print new frame with cube and TODO: axis
        txt_out.write(str(frame_num) + ' %f %f %f\n' % (yaw_predicted, pitch_predicted, roll_predicted))
        utils.plot_pose_cube(frame, yaw_predicted, pitch_predicted, roll_predicted, (x_min + x_max) / 2, (y_min + y_max) / 2, size = 200)
        # Plot expanded bounding box
        cv2.rectangle(frame, (x_min, y_min), (x_max, y_max), (0,255,0), 3)
        out.write(frame)
                yaw_predicted = F.softmax(yaw)
                pitch_predicted = F.softmax(pitch)
                roll_predicted = F.softmax(roll)
                # Get continuous predictions in degrees.
                yaw_predicted = torch.sum(yaw_predicted.data[0] * idx_tensor) * 3 - 99
                pitch_predicted = torch.sum(pitch_predicted.data[0] * idx_tensor) * 3 - 99
                roll_predicted = torch.sum(roll_predicted.data[0] * idx_tensor) * 3 - 99
                # Print new frame with cube and axis
                txt_out.write(str(frame_num) + ' %f %f %f\n' % (yaw_predicted, pitch_predicted, roll_predicted))
                # utils.plot_pose_cube(frame, yaw_predicted, pitch_predicted, roll_predicted, (x_min + x_max) / 2, (y_min + y_max) / 2, size = bbox_width)
                utils.draw_axis(frame, yaw_predicted, pitch_predicted, roll_predicted, tdx = (x_min + x_max) / 2, tdy= (y_min + y_max) / 2, size = bbox_height/2)
                # Plot expanded bounding box
                # cv2.rectangle(frame, (x_min, y_min), (x_max, y_max), (0,255,0), 1)

        frame_num += 1
            # Peek next frame detection
            next_frame_num = int(bbox_line_list[idx+1].strip('\n').split(' ')[0])
            # print 'next_frame_num ', next_frame_num
            if next_frame_num == det_frame_num:
                idx += 1
                line = bbox_line_list[idx].strip('\n').split(' ')
                det_frame_num = int(line[0])
            else:
                break

    while True:
        ret, frame = video.read()
        if ret == False:
            out.release()
            video.release()
            bbox_file.close()
            txt_out.close()
            sys.exit(0)
        idx += 1
        out.write(frame)
        frame_num += 1

    out.release()
    video.release()
    txt_out.close()