The Cityscapes Dataset for Semantic Urban Scene Understanding SUPPLEMENTAL MATERIAL

Transcription

1 The Cityscapes Dataset for Semantic Urban Scene Understanding SUPPLEMENTAL MATERIAL Marius Cordts 1,2 Mohamed Omran 3 Sebastian Ramos 1,4 Timo Rehfeld 1,2 Markus Enzweiler 1 Rodrigo Benenson 3 Uwe Franke 1 Stefan Roth 2 Bernt Schiele 3 1 Daimler AG R&D, 2 TU Darmstadt, 3 MPI Informatics, 4 TU Dresden A. Related Datasets In Tab. 7 we provide a comparison to other related datasets in terms of the type of annotations, the meta information provided, the camera perspective, the type of scenes, and their size. The selected datasets are either of large scale or focus on street scenes. B. Class Definitions Table 8 provides precise definitions of our annotated classes. These definitions were used to guide our labeling process, as well as quality control. In addition, we include a typical example for each class. The annotators were instructed to make use of the depth ordering and occlusions of the scene to accelerate labeling, analogously to LabelMe [59]; see Fig. 6 for an example. In doing so, distant objects are annotated first, while occluded parts are annotated with a coarser, conservative boundary (possibly larger than the actual object). Subsequently, the occluder is annotated with a polygon that lies in front of the occluded part. Thus, the boundary between these objects is shared and consistent. Holes in an object through which a background region can be seen are considered to be part of the object. This allows keeping the labeling effort within reasonable bounds such that objects can be described via simple polygons forming simply-connected sets. C. Example Annotations Figure 7 presents several examples of annotated frames from our dataset that exemplify its diversity and difficulty. All examples are taken from the and val splits and were chosen by searching for the extremes in terms of the number of traffic participant instances in the scene; see Fig. 7 for details. Figure 6. Exemplary labeling process. Distant objects are annotated first and subsequently their occluders. This ensures the boundary between these objects to be shared and consistent. D. Detailed Results In this section, we present additional details regarding our control experiments and baselines. Specifically, we give individual class scores that complement the aggregated scores in the main paper. Moreover, we provide details on the ing procedure for all baselines. Finally, we show additional qualitative results of all methods. D.1. Semantic labeling Tables 9 and 11 list all individual class-level IoU scores for all control experiments and baselines. Tables 10 and 12 give the corresponding instance-normalized iiou scores. In addition, Figs. 8 and 9 contain qualitative examples of these methods. Basic setup. All baselines relied on single frame, monocular LDR images and were preed on ImageNet [58], i.e. their underlying CNN was generally initialized with ImageNet VGG weights [67]. Subsequently, the CNNs were finetuned on Cityscapes using the respective portions listed in Tab. 4. In our own FCN [40] experiments, we additionally investigated first preing on PASCAL-Context [44], but found this to not influence performance given a sufficiently large number of ing iterations. Most baselines applied a subsampling of the input image, c.f. Tab. 4, probai

2 Dataset Labels Color Video Depth Camera Scene #images #classes [58] B Mixed Mixed 150 k 1000 [13] B, C Mixed Mixed 20 k (B), 10 k (C) 20 [44] D Mixed Mixed 20 k 400 [37] C Mixed Mixed 300 k 80 [68] D, C Kinect Pedestrian Indoor 10 k 37 [18] B, D a Laser, Stereo Car Suburban 15 k (B), 700 (D) 3 (B), 8 (D) [6] D Car Urban [34] D Stereo, Manual Car Urban 70 7 [60] D Stereo Car Urban [2] D Pedestrian Urban [64] C Stereo Car Facades [55] D 3D mesh Pedestrian Urban [74] D Laser Car Suburban 400 k 27 Ours D, C Stereo Car Urban 5 k (D), 20 k (C) 30 a Including the annotations of 3 rd party groups [21, 28, 31, 32, 57, 63, 76, 79] Table 7. Comparison to related datasets. We list the type of labels provided, i.e. object bounding boxes (B), dense pixel-level semantic labels (D), coarse labels (C) that do not aim to label the whole image. Further, we mark if color, video, and depth information are available. We list the camera perspective, the scene type, the number of images, and the number of semantic classes. bly due to time or memory consts. Only Adelaide [36], Dilated10 [78], and our FCN experiments were conducted on the full-resolution images. In the first case, a new random patch of size pixels was drawn at each iteration. In our FCN ing, we split each image into two halves (left and right) with an overlap that is sufficiently large considering the network s receptive field. Own baselines. The ing procedure of all our FCN experiments follows [40]. We use three-stage ing with subsequently smaller strides, i.e. first FCN-32s, then FCN- 16s, and then FCN-8s, always initializing with the parameters from the previous stage. We add a 4 th stage for which we reduce the learning rate by a factor of 10. The ing parameters are identical to those publicly available for ing on PASCAL-Context [44], except that we reduce the learning rate to account for the increased image resolution. Each stage is ed until convergence on the validation set; pixels with void ground truth are ignored such that they do not induce any gradient. Eventually, we re on and val together with the same number of epochs, yielding , , , and 5950 iterations for stages 1 through 4. Note that each iteration corresponds to half of an image (see above). For the variant with factor 2 downsampling, no image splitting is necessary, yielding , , , and 5950 iterations in the respective stages. The variant only ed on val (full resolution) uses for validation, leading to , , , and 0 iterations in the 4 stages. Our last FCN variant is ed using the coarse annotations only, with , , , and 0 iterations in the respective stage; pixels with void ground truth are ignored here as well. 3 rd -party baselines. Note that for the following descriptions of the 3 rd -party baselines, we have to rely on authorprovided information. SegNet [3] ing for both the basic and extended variant was performed until convergence, yielding approximately 50 epochs. Inference takes 0.12 s per image. DPN [39] was ed using the original procedure, while using all available Cityscapes annotations. For ing CRF as RNN [80], an FCN-32s model was ed for 3 days on using a GPU. Subsequently an FCN-8s model was ed for 2 days, and eventually the model was further finetuned including the CRF-RNN layers. Testing takes 0.7 s on half-resolution images. For ing DeepLab on the fine annotations, denoted DeepLab-LargeFOV-Strong, the authors applied the ing procedure from [8]. The model was ed on for iterations until convergence on val. Then val was included in the ing set for another iterations. In both cases, a mini-batch size of 10 was applied. Each ing iteration lasts 0.5 s, while inference including the dense CRF takes 4 s per image. The DeepLab variant including our coarse annotations, termed DeepLab- LargeFOV-StrongWeak, followed the protocol in [47] and is initialized from the DeepLab-LargeFOV-Strong model. Each mini-batch consists of 5 finely and 5 coarsely annotated images and ing is performed for iterations until convergence on val. Then, ing was continued for another iterations on and val. Adelaide [36] was ed for 8 days using random crops of the input image as described above. Inference on a single image takes 35 s. The best performing baseline, Dilated10 [78], is a convolutional network that consists of a front-end prediction module and a context aggregation module. The front-end module is an adaptation of the VGG-16 network based on dilated convolutions. The context module uses dilated convolutions ii

3 to systematically expand the receptive field and aggregate contextual information. This module is derived from the Basic" network, where each layer has C = 19 feature maps. The total number of layers in the context module is 10, hence the name Dilation10. The increased number of layers in the context module (10 for Cityscapes versus 8 for PASCAL VOC) is due to the higher input resolution. The complete Dilation10 model is a pure convolutional network: there is no CRF and no structured prediction. The Dilation10 network was ed in three stages. First, the frontend prediction module was ed for iterations on randomly sampled crops of size , with learning rate 10 4, momentum 0.99, and batch size 8. Second, the context module was ed for iterations on whole (uncropped) images, with learning rate 10 4, momentum 0.99, and batch size 100. Third, the complete model (front-end + context) was jointly ed for iterations on halves of images (input size , including padding), with learning rate 10 5, momentum 0.99, and batch size 1. D.2. Instance-level semantic labeling For our instance-level semantic labeling baselines and control experiments, we rely on Fast R-CNN [19] and proposal regions from either MCG (Multiscale Combinatorial Grouping [1]) or from the ground truth annotations. We use the standard ing and testing parameters for Fast R-CNN. Training starts with a model pre-ed on ImageNet [58]. We use a learning rate of and stop when the validation error plateaus after iterations. At test time, one score per class is assigned to each object proposal. Subsequently, thresholding and non-maximum suppression is applied and either the bounding boxes, the original proposal regions or their convex hull are used to generate the predicted masks of each instance. Quantitative results of all classes can be found in Tables 13 to 16 and qualitative results in Fig. 12. iii

4 Category Class Definition Examples human person 1 All humans that would primarily rely on their legs to move if necessary. Consequently, this label includes people who are standing/sitting, or otherwise stationary. This class also includes babies, people pushing a bicycle, or standing next to it with both legs on the same side of the bicycle. rider 1 Humans relying on some device for movement. This includes drivers, passengers, or riders of bicycles, motorcycles, scooters, skateboards, horses, Segways, (inline) skates, wheelchairs, road cleaning cars, or convertibles. Note that a visible driver of a closed car can only be seen through the window. Since holes are considered part of the surrounding object, the human is included in the car label. vehicle car 1 tinuous body shape (i.e. the driver s cabin and cargo compartment are one). Does not include This includes cars, jeeps, SUVs, vans with a con- trailers, which have their own separate class. truck 1 This includes trucks, vans with a body that is separate from the driver s cabin, pickup trucks, as well as their trailers. bus 1 This includes buses that are intended for 9+ persons for public or long-distance transport. 1 All vehicles that move on rails, e.g. trams, s. 1 Single instance annotation available. 2 Not included in challenges. Table 8. List of annotated classes including their definition and typical example. iv

5 Category Class Definition Examples vehicle motorcycle 1 without the driver or other passengers. The latter This includes motorcycles, mopeds, and scooters receive the label rider. bicycle 1 This includes bicycles without the cyclist or other passengers. The latter receive the label rider. caravan 1,2 Vehicles that (appear to) contain living quarters. This also includes trailers that are used for living and has priority over the trailer class. trailer 1,2 Includes trailers that can be attached to any vehicle, but excludes trailers attached to trucks. The latter are included in the truck label. nature vegetation Trees, hedges, and all kinds of vertically growing vegetation. Plants attached to buildings/walls/fences are not annotated separately, and receive the same label as the surface they are supported by. terrain Grass, all kinds of horizontally spreading vegetation, soil, or sand. These are areas that are not meant to be driven on. This label may also include a possibly adjacent curb. Single grass stalks or very small patches of grass are not annotated separately and thus are assigned to the label of the region they are growing on. 1 Single instance annotation available. 2 Not included in challenges. Table 8. List of annotated classes including their definition and typical example. (continued) v

6 Category Class Definition Examples construction building Includes structures that house/shelter humans, e.g. low-rises, skyscrapers, bus stops, car ports. Translucent buildings made of glass still receive the label building. Also includes scaffolding attached to buildings. wall Individually standing walls that separate two (or more) outdoor areas, and do not provide support for a building. fence Structures with holes that separate two (or more) outdoor areas, sometimes temporary. guard rail 2 Metal structure located on the side of the road to prevent serious accidents. Rare in inner cities, but occur sometimes in curves. Includes the bars holding the rails. bridge 2 Bridges (on which the ego-vehicle is not driving) including everything (fences, guard rails) permanently attached to them. tunnel 2 Tunnel walls and the (typically dark) space encased by the tunnel, but excluding vehicles. 1 Single instance annotation available. 2 Not included in challenges. Table 8. List of annotated classes including their definition and typical example. (continued) vi

7 Category Class Definition Examples object traffic sign Front part of signs installed by the state/city authority with the purpose of conveying information to drivers/cyclists/pedestrians, e.g. traffic signs, parking signs, direction signs, or warning reflector posts. traffic light The traffic light box without its poles in all orientations and for all types of traffic participants, e.g. regular traffic light, bus traffic light, traffic light. pole Small, mainly vertically oriented poles, e.g. sign poles or traffic light poles. This does not include objects mounted on the pole, which have a larger diameter than the pole itself (e.g. most street lights). pole group 2 Multiple poles that are cumbersome to label individually, but where the background can be seen in their gaps. sky sky Open sky (without tree branches/leaves) 1 Single instance annotation available. 2 Not included in challenges. Table 8. List of annotated classes including their definition and typical example. (continued) vii

8 Category Class Definition Examples flat road Horizontal surfaces on which cars usually drive, including road markings. Typically delimited by curbs, rail tracks, or parking areas. However, road is not delimited by road markings and thus may include bicycle lanes or roundabouts. sidewalk Horizontal surfaces designated for pedestrians or cyclists. Delimited from the road by some obstacle, e.g. curbs or poles (might be small), but not only by markings. Often elevated compared to the road and often located at the side of a road. The curbs are included in the sidewalk label. Also includes the walkable part of traffic islands, as well as pedestrian-only zones, where cars are not allowed to drive during regular business hours. If it s an all-day mixed pedestrian/car area, the correct label is ground. parking 2 Horizontal surfaces that are intended for parking and separated from the road, either via elevation or via a different texture/material, but not separated merely by markings. rail track 2 Horizontal surfaces on which only rail cars can normally drive. If rail tracks for trams are embedded in a standard road, they are included in the road label. 1 Single instance annotation available. 2 Not included in challenges. Table 8. List of annotated classes including their definition and typical example. (continued) viii

9 Category Class Definition Examples void ground 2 All other forms of horizontal ground-level structures that do not match any of the above, for example mixed zones (cars and pedestrians), roundabouts that are flat but delimited from the road by a curb, or in general a fallback label for horizontal surfaces that are difficult to classify, e.g. due to having a dual purpose. dynamic 2 Movable objects that do not correspond to any of the other non-void categories and might not be in the same position in the next day/hour/minute, e.g. movable trash bins, buggies, luggage, animals, chairs, or tables. static 2 This includes areas of the image that are difficult to identify/label due to occlusion/distance, as well as non-movable objects that do not match any of the non-void categories, e.g. mountains, street lights, reverse sides of traffic signs, or permanently mounted commercial signs. ego vehicle 2 unlabeled 2 Since a part of the vehicle from which our data was recorded is visible in all frames, it is assigned to this special label. This label is also available at test time. Pixels that were not explicitly assigned to a label. out of roi 2 rectification border 2 1 Single instance annotation available. 2 Not included in challenges. Narrow strip of 5 pixels along the image borders that is not considered for ing or evaluation. This label is also available at test-time. Areas close to the image border that contain artifacts resulting from the stereo pair rectification. This label is also available at test time. Table 8. List of annotated classes including their definition and typical example. (continued) ix

10 Largest number of instances and persons Largest number of riders Largest number of cars Largest number of bicycles Largest number of buses Largest number of trucks Largest number of motorcycles Large spatial variation of persons Fewest number of instances Figure 7. Examples of our annotations on various images of our and val sets. The images were selected based on criteria overlayed on each image. x

11 road sidewalk building wall fence pole traffic light traffic sign vegetation terrain sky person rider car truck bus motorcycle bicycle mean IoU static fine (SF) static coarse (SC) GT segmentation with SF GT segmentation with SC GT segmentation with [40] GT subsampled by GT subsampled by GT subsampled by GT subsampled by GT subsampled by GT subsampled by GT subsampled by nearest ing neighbor Table 9. Detailed results of our control experiments for the pixel-level semantic labeling task in terms of the IoU score on the class level. All numbers are given in percent. See the main paper for details on the listed methods. person rider car truck bus motorcycle bicycle mean iiou static fine (SF) static coarse (SC) GT segmentation with SF GT segmentation with SC GT segmentation with [40] GT subsampled by GT subsampled by GT subsampled by GT subsampled by GT subsampled by GT subsampled by GT subsampled by nearest ing neighbor Table 10. Detailed results of our control experiments for the pixel-level semantic labeling task in terms of the instance-normalized iiou score on the class level. All numbers are given in percent. See the main paper for details on the listed methods. xi

12 val coarse sub road sidewalk building wall fence pole traffic light traffic sign vegetation terrain sky person rider car truck bus motorcycle bicycle mean IoU FCN-32s FCN-16s FCN-8s FCN-8s FCN-8s FCN-8s [3] ext [3] basic [39] [80] [8] [47] [36] [78] Table 11. Detailed results of our baseline experiments for the pixel-level semantic labeling task in terms of the IoU score on the class level. All numbers are given in percent and we indicate the used ing data for each method, i.e. fine, val fine, coarse extra, as well as a potential downscaling factor (sub) of the input image. See the main paper and Appendix D.1 for details on the listed methods. val coarse sub person rider car truck bus motorcycle bicycle mean iiou FCN-32s FCN-16s FCN-8s FCN-8s FCN-8s FCN-8s [3] extended [3] basic [39] [80] [8] [47] [36] [78] Table 12. Detailed results of our baseline experiments for the pixel-level semantic labeling task in terms of the instance-normalized iiou score on the class level. All numbers are given in percent and we indicate the used ing data for each method, i.e. fine, val fine, coarse extra, as well as a potential downscaling factor (sub) of the input image. See the main paper and Appendix D.1 for details on the listed methods. Proposals Classifier person rider car MCG regions FRCN MCG bboxes FRCN MCG hulls FRCN GT bboxes FRCN GT regions FRCN MCG regions GT MCG bboxes GT MCG hulls GT Table 13. Detailed results of our baseline experiments for the instance-level semantic labeling task in terms of the region-level average precision scores AP on the class level. All numbers are given in percent. See the main paper and Appendix D.2 for details on the listed methods. truck bus motorcycle bicycle mean AP xii

13 Proposals Classifier person rider car MCG regions FRCN MCG bboxes FRCN MCG hulls FRCN GT bboxes FRCN GT regions FRCN MCG regions GT MCG bboxes GT MCG hulls GT Table 14. Detailed results of our baseline experiments for the instance-level semantic labeling task in terms of the region-level average precision scores AP 50% for an overlap value of 50 %. All numbers are given in percent. See the main paper and Appendix D.2 for details on the listed methods. truck bus motorcycle bicycle mean AP 50% Proposals Classifier person rider car MCG regions FRCN MCG bboxes FRCN MCG hulls FRCN GT bboxes FRCN GT regions FRCN MCG regions GT MCG bboxes GT MCG hulls GT Table 15. Detailed results of our baseline experiments for the instance-level semantic labeling task in terms of the region-level average precision scores AP 100m for objects within 100 m. All numbers are given in percent. See the main paper and Appendix D.2 for details on the listed methods. truck bus motorcycle bicycle mean AP 100m Proposals Classifier person rider car MCG regions FRCN MCG bboxes FRCN MCG hulls FRCN GT bboxes FRCN GT regions FRCN MCG regions GT MCG bboxes GT MCG hulls GT Table 16. Detailed results of our baseline experiments for the instance-level semantic labeling task in terms of the region-level average precision scores AP 50m for objects within 50 m. All numbers are given in percent. See the main paper and Appendix D.2 for details on the listed methods. truck bus motorcycle bicycle mean AP 50m xiii

14 Image Annotation static fine (SF) static coarse (SC) GT segmentation w/ SF GT segmentation w/ SC GT segmentation w/ [40] GT subsampled by 2 GT subsampled by 8 GT subsampled by 32 GT subsampled by 128 nearest ing neighbor Figure 8. Exemplary output of our control experiments for the pixel-level semantic labeling task, see the main paper for details. The image is part of our test set and has both, the largest number of instances and persons. xiv

15 Image Annotation FCN-32s FCN-8s FCN-8s half resolution FCN-8s ed on coarse SegNet basic [39] DPN [8] CRF as RNN [3] DeepLab LargeFOV StrongWeak [47] Adelaide [36] Dilated10 [78] Figure 9. Exemplary output of our baselines for the pixel-level semantic labeling task, see the main paper for details. The image is part of our test set and has both, the largest number of instances and persons. xv

16 Image Annotation static fine (SF) static coarse (SC) GT segmentation w/ SF GT segmentation w/ SC GT segmentation w/ [40] GT subsampled by 2 GT subsampled by 8 GT subsampled by 32 GT subsampled by 128 nearest ing neighbor Figure 10. Exemplary output of our control experiments for the pixel-level semantic labeling task, see the main paper for details. The image is part of our test set and has the largest number of car instances. xvi

17 Image Annotation FCN-32s FCN-8s FCN-8s half resolution FCN-8s ed on coarse SegNet basic [39] DPN [8] CRF as RNN [3] DeepLab LargeFOV StrongWeak [47] Adelaide [36] Dilated10 [78] Figure 11. Exemplary output of our baseline experiments for the pixel-level semantic labeling task, see the main paper for details. The image is part of our test set and has the largest number of car instances. xvii

18 Largest number of instances and persons Annotation FRCN + MCG bboxes FRCN + MCG regions FRCN + GT bboxes FRCN + GT regions Largest number of cars Annotation FRCN + MCG bboxes FRCN + MCG regions FRCN + GT bboxes FRCN + GT regions Figure 12. Exemplary output of our control experiments and baselines for the instance-level semantic labeling task, see the main paper for details. xviii