I am trying to test the system in outdoor environments using different datasets and configurations. Below is a detailed breakdown of the methods I used:
KITTI Odometry (Processed using TUM-like parser and TUM-style groundtruth data)
KITTI Odometry (TUM-style groundtruth data processed using the TUM parser, combined with LiDAR-based depth maps)
KITTI Odometry (Original KITTI data format processed with pykitti tool and my custom parser, without depth images)
KITTI Odometry (Same approach as above, but depth maps were generated from projected Velodyne point clouds)
KITTI Raw Dataset (Two different subsets processed using pykitti, no depth images included)
KITTI Depth Dataset (Processed using only RGB images and my custom parser)
NCLT Dataset (Encountered an error indicating "insufficient motion"; system could not be run successfully, retry needed)
For each method listed above, the system was run at least once. However, I consistently encountered significant errors in both trajectory estimation and reconstruction across all setups.
Possible Causes of the Errors:
Groundtruth Issues: I tried both the original KITTI groundtruth format and the TUM-style version. Despite this, groundtruth processing may still be causing errors.
Optimization Parameters: The optimization settings were tuned for KITTI's original data, and additional improvements were applied. However, the system appears to work better with indoor datasets, suggesting that factors like sunlight or outdoor conditions might be contributing to the issues.
Frame Processing Delays: There is a notable difference in the frame processing times between TUM and KITTI datasets. For instance, processing 20 frames takes 0.6 seconds for TUM, whereas for KITTI, the same amount of frames takes 1.9 seconds.This delay might be leading to skipped frames and insufficient time for accurate reconstruction.
If anyone has any ideas or suggestions to help resolve these issues, your input would be greatly appreciated!
I am trying to test the system in outdoor environments using different datasets and configurations. Below is a detailed breakdown of the methods I used:
For each method listed above, the system was run at least once. However, I consistently encountered significant errors in both trajectory estimation and reconstruction across all setups. Possible Causes of the Errors:
Groundtruth Issues: I tried both the original KITTI groundtruth format and the TUM-style version. Despite this, groundtruth processing may still be causing errors.
Optimization Parameters: The optimization settings were tuned for KITTI's original data, and additional improvements were applied. However, the system appears to work better with indoor datasets, suggesting that factors like sunlight or outdoor conditions might be contributing to the issues.
Frame Processing Delays: There is a notable difference in the frame processing times between TUM and KITTI datasets. For instance, processing 20 frames takes 0.6 seconds for TUM, whereas for KITTI, the same amount of frames takes 1.9 seconds.This delay might be leading to skipped frames and insufficient time for accurate reconstruction.
If anyone has any ideas or suggestions to help resolve these issues, your input would be greatly appreciated!