IDD Comprehension

1
IDD Comprehension
2
Existing Indian Driving Datasets

• Detection Dataset
• IDD Segmentation
• IDD Temporal Train
• Fine-Grained Vehicle Detection
• IDD 3D Dataset

3
List of Datasets for Data Capture
S.No Dataset Name Purpose Task to be performed
1 Driving Place Recognition Driving Place Recognition datasets are designed to aid research and development in autonomous driving and computer vision tasks related to place recognition. Condition-Invariant Place Recognition. Viewpoint-Invariant Place Recognition.  
2 Pedestrian Crosswalk Behaviour The Pedestrian Crosswalk Behaviour dataset is likely to be a collection of visual data captured from CCTV cameras or dashcams positioned near pedestrian crosswalks Pedestrian Detection Pedestrian Crosswalk Behaviour
3 Trajectory Prediction Trajectory Prediction datasets are collections of data used to aid research and development in the field of trajectory prediction. Pedestrian Trajectory Prediction Near Accident Prediction Pedestrian Intention Estimation
4
Infrastructure Requirements For Driving
Place Recognition
5
Additional Material Required for Data Capture
Sensor Quantity Resolution Configuration Manufacturer/Model
LiDAR 1 64 channels (vertical) 1024 channel (horizontal) 10 Hz capture. XYZ, Intensity, Reflectivity, Range Ouster OS1 sensor
Cameras 6 2048 x 1536 BayerRG8 format 10 Hz capture FLIR Blackfly S, C-mount
Lens 6 - UC Series Fixed focal length 12/25mm Edmund optics
GPS 1 - G-Star IV BU-353-S4 sensor 1Hz GlobalSat

• For data capture, we can onboard the vehicle with
  existing available sensors/equipment to collect
  the necessary data.

6
Sensors and setup

• 1 front stereo camera or 2 different viewpoints
  cameras.
• 2 left, right cameras
• 1 back camera or 2 different viewpoints cameras.
• 2D and 3D lidar to be considered for 3D VPR
• 1 GPS system
• INS sensor
• How do our cameras compare to SOTA datasets, like
  oxford car cameras?

7
Memory Allocation

• The data generated per minute is approximately
  15GB.
• The data volume is adequate to occupy a 1TB disk
  during approximately one hour of continuous data
  collection.
• To ensure optimal write speeds to the disk, we
  have implemented the usage of SSDs with a high
  transfer rate. The data is directly recorded onto
  the external disk, effectively eliminating any
  additional data transfer overhead on the system.
• To provide an added layer of safety, an
  additional 5 TB of SSD storage is recommended.
  This will help accommodate any potential increase
  in data generation and ensure ample space for
  future requirements, further enhancing the
  reliability and efficiency of the data storage
  system.

Project Memory Allocation Memory Used Available Memory Additional Storage
IDD-Comprehension 50TB - - 5TB
8
Annotaion
9
Estimated Time To Start Data Capture

• The time it takes to initiate data capture is
  contingent upon the prevailing climatic
  conditions.

10
Issues to be addressed before and in the early
attempts of data structure.

• I will attache my comments in the following
  slides and then you incorporate them in this file
  accordingly.

11
Data collection and synchronization

• The data format of IDD-3D which is used by the
  current hardware setup to be compared with the
  SOTA dataset format.
• Ensure the synchronization between all cameras
  and the LIDAR and GPS.
• Using the same setup for capturing under the rain
  or using different camera (dashcam).

12
Reference Reading list

1. Visual Place Recognition A Tutorial
2. Dont Look Back Robustifying Place
   Categorization for Viewpoint- and
   Condition-Invariant Place Recognition
3. General Place Recognition Survey Towards the
   Real-world Autonomy Age.
4. 1 Year, 1000km The Oxford RobotCar Dataset
5. Real-time Kinematic Ground Truth for the Oxford
   RobotCar Dataset.
6. A Survey on Deep Visual Place Recognition
7. Visual place recognition A survey from deep
   learning perspective