The KITTI-ROAD Evaluation Benchmark for Road Detection Algorithms
08.06.2014 Jannik Fritsch
Honda Research Institute Europe, Offenbach, Germany
Presented material created together with Tobias Kuehnl
Research Institute for Cognition and Robotics, Bielefeld University, Bielefeld, Germany
Contents
• The KITTI ROAD Dataset
• Classical Pixel/Cell-based Performance Measures
• Cell-based Evaluation in Metric Space
• Proposal for Behavior-level Evaluation Benchmark
• Behavior-based Evaluation
• Short Discussion of Current Results on Website
• Disadvantages of KITT-ROAD data set
• Summary
KITTI – Collection of Several Vision Benchmarks for
Automotive Domain
KITTI-ROAD – A New Benchmark for Road Detection Evaluation
KITTI – ROAD Dataset
• Derived from the KITTI dataset stereo, velodyne, and IMU data available!
• Separated in 289 training (with annotation) and 290 test images
• 3 types of city roads
• Annotations for road-area and ego-lane (only UM)
Example used in this talk:
Evaluation of road-area on complete URBAN dataset and ego-lane on UM_LANE UU_ROAD: Urban Unmarked Road
UM_ROAD: Urban Marked Road
UMM_ROAD: Urban Multiple Marked Road
UM_LANE: EgoLane on UM_ROAD
Evaluation Criteria for Road Terrain Detection
• Pixel-level correctness
– classical measure
Kang et al. 2011, Alvarez et al. 2008, Wu et al. 2011
• Boundary position and deviation
– (μ,σ) at one/multiple distances
Zhao et al. 2012, Kuehnl et al. 2011, et al. Serfling 2008
• Unoccupied lane length
– distance d up to obstacle
Gumpp et al 2011
• Corridor width
– continuous (μ,σ) or discrete classes
Kuehnl et al 2012
• Model-based lane shape similarity
– clothoid parameter deviations
Gopalan et al. 2012, Gackstaetter 2011
Perspective BEV
αααα αααα
Classical Evaluation metric
Proposed new evaluation metric -- abstraction from
exact boundary
Application- Relevant Space
Classical Pixel/Cell-based Performance Measures
• Pixel-based evaluation measures employed typically:
• Ideal performance: Set threshold TH to value maximizing harmonic F-measure (beta=1, i.e. equal weight to Precision and Recall)
• Average performance: Average Precision (AP) used in PASCAL VOC evaluations
All measures are applicable to both: 1) Perspective image and 2) Metric BEV space
Comparison Algorithms
1. Baseline (BL):
– Accumulating ground truth annotations provides probabilistic spatial prior
2. Here: Geometric Context (GC) [Hoiem2007
]: (somewhat unfair, as it detects any planar surface) – Segment into Superpixels and calculate probability distribution of surface orientations3. Online: SPatial RAY features (SPRAY) [Kuehnl2012]:
– Separate classification of appearance and space
4. Online: Convolutional Neural Network (CNN) [Alvarez2012]:
– Combination of offline and online methods
Perspective Bird’s-Eye-View (BEV)
road-area ego-lane
road-area
ego-lane
Image Area for Pixel-based Evaluation
• Metric BEV area
– covers +/-10m x 40m
• Perspective image
– Covers only area matching to BEV (red polygon):
same metric area evaluated also in Perspective space BUT different weighting due to different number of pixels (see pink rectangle)
– Perspective: Near range has high influence, Far range has low influence
– BEV: Near and far range have same influence
• Note:
– Here both Metric results and Perspective results are provided for comparison – KITTI-ROAD benchmark on webserver evaluates BEV only
Perspective BEV
-10 0 10
46
26
6
Results for Road-area on URBAN Dataset
• BL surprisingly good due to
– Low traffic density = road usually free
– Strong lighting variations = hard for vision algorithms
• Differences between algorithms more pronounced in BEV Results emphasize importance of BEV evaluation
Perspective Bird’s-Eye-View (BEV) Perspective
BEV
Results for Ego-lane on UM Dataset
• Extremely high accuracy due to many TNs (ego-lane covers only small part in BEV)
... but similar precision values as for road-area
• Again: BL extremely good
• SPRAY approach only better in BEV evaluation
Perspective Bird’s-Eye-View (BEV) Perspective
BEV
Basic Idea of Behavior-based Evaluation
• Single track model to predict vehicle motion
• Use different steering angles for generating corridor hypotheses
• Obtain fitness value for each hypothesis by integrating
ego-lane confidences covered by corridor area (2.2m width = vehicle width)
• Select best corridor hypothesis to represent driveable ego-lane Corridor representation abstracts from pixel-level ego-lane area
0.71 0.08
SPRAY
Ego-lane confidences
Corridor hypotheses
Corridor fitness
… …
Best corridor hypotheses
Examples of Corridor Hypotheses on KITTI ROAD benchmark
Note:
Most of the time the road is straight (exception: rural roads)
Only small amount of recorded data/benchmark contains curvy scenes
Characteristics of Resulting Corridor Hypothesis
Different applications of parametric corridor possible:
– LKAS support on roads with no/bad lane markings
– Detection of too narrow passages independent from road boundary type – ….
• Evaluation issue: Corridor covers only subset of ego-lane ground truth area Classical evaluation metric not fully applicable (large number of irrelevant FN)
Perspective BEV
Generate corridor hypothesis
Evaluation
Evaluation for Ego-lane Corridor Hypothesis
• Lateral quality based on precision at different distances
• Longitudinal quality obtained by “shrinking” corridor to single line (avoid high FN)
– TP only if sufficient overlap in original pixel representation (2.0m out of 2.2m corridor) – Evaluation of F1 measure and Hitrate (successful match of hypothesis to ground truth)
Proposal for new performance measure for behavior-based performance
(But: evaluation on single image, ADAS needs temporal integration events)
Individual Results (UU,UM,UMM) Available Online
Individual Results (UU,UM,UMM) Available Online
Some Thoughts on the Current Results
Choice of sensor data used
• Mono: Only few submissions, partially good results (SPRAY, but needs ext. calib.)
• Stereo: very powerful but not better than mono (if ignoring color/texture)
– Planar depth information is not discriminative (SP)
• roads are not completely flat (small bending for rain drainage!)
• flat gras next to road
• gravel & tram rail area
• low curbs
• sidewalk at same height but from different material
• large distances are generally challenging (especially for stereo) – Combination with BL helps to avoid some pitfalls (SP+BL)
– Completely different use of stereo data leads to similar results (RES3D-Stereo)
• Stereo with Color/Texture
– Top position exceeding best mono (ProbBoost)
– Diverse results at lower positions (HistonBoost, ANN)
• Velodyne: results much better than pure stereo (RES3D-Velo)
– But not superior to mono emphasizes importance of color/texture – Better depth information in large distances
Disadvantages of KITTI-ROAD benchmark
• Currently only small amount of data (300/300), but total recordings available only allow for limited expansion of data set
• KITTI captures only German city roads and a little bit of rural roads very restricted form of roads (US: wider, JP: city roads smaller, …)
• Exposure control is not perfect, especially the road area is sometimes over-exposed Appearance changes are probably more drastic than in modern cameras
• Straight roads are common and enable high performance of simple spatial prior (BL) BUT: the interesting situation are curves
more curve situations needed, difficult with KITTI since most roads are straight (not enough data to create “Curve” benchmark)
• Currently only polygonal area annotation, so only “implicit” boundaries
could be added, if common understanding of boundary nature could be reached
(see also paper in next session after the coffee break)
Summary
• A novel benchmark: KITTI ROAD dataset with 3 urban categories
• Online evaluation on webserver:
– Pixel-based road-area in BEV space – Pixel-based ego-lane in BEV space
– Behavior-based ego-lane corridor hypothesis