July 2020
tl;dr: Class rebalance of minority helps in object detection for nuscenes dataset.
The class balanced sampling and class-grouped heads are useful to handle imbalanced object detection.
- DS sampling:
- increases sample density of rare classes to avoid gradient vanishing
- count instances and samples (frames). Resample so that samples for each class is on the same order of magnitude.
- Class balanced grouping: each group has a separate head.
- Classes of similar shapes or sizes should be grouped.
- Instance numbers of diff groups should be balanced properly.
- Supergroups:
- cars (majority classes)
- truck, construction vehicle
- bus, trailer
- barrier
- motorcycle, bicycle
- pedestrian, traffic cone
- Fit ground plane and plant GT back in.
- Bag of tricks
- Accumulate 10 frames (0.5 seconds) to form a dense lidar BEV
- AdamW + One cycle policy
- Regress vx and vy. If bicycle speed is above a certain thresh, then it is with rider.
- Questions and notes on how to improve/revise the current work