April 2020
tl;dr: A boosted version of Fast and Furious that uses map information.
IntentNet is heavily inspired by Fast and Furious (also by Uber ATG). Both combines perception, tracking and prediction by generating bbox with waypoints. In comparison, IntentNet extends the horizon from 1s to 3s, predicts discrete high level behaviors, and uses map information. Note that both IntentNet and Fast and Furious do not perform motion planning.
The concatenation of images may introduce more compuation than LSTM, but itself does not preclude real-time performance. Fast and Furious also uses concatenation of input and achieves real-time performance. It is further extended to neural motion planner.
IntentNet somewhat inspired MultiPath but IntentNet's loss only predict one path instead of one path per intent, making IntentNet unsuitable for multimodal prediction.
- Map info rendered as 17-channel binary BEV map.
- Architecture: Late fusion of map and lidar (first 3 blocks split, with the last block concatenated)
- Concatenate past 10 frames along channel dimension --> cannot recycle computation. Why don't use LSTM for better real-time performance?
- Predict discrete intention (8 classes)
- Predicts current location/detection
- Predicts future waypoint.
- Loss (FaF-like, but with time discount)
- YOLO-like, anchor-based object classification (binary focal loss). .
- Trajectory regression of waypoint positions, orientation, and w and h (constant over time) (smooth L2 loss)
- Loss is discounted by 0.97 each timestep into the future
- 144 x 80 meters (0.2 m per pixel), 720 x 400 pixels.
- The paper argues that using intermediate perception and tracking results for intent prediction restricts the information and may have suboptimal performance. --> This may be true in theory but in practice how much effect the image information has to intent prediction is questionable. The ultimate question is, can human being figure out the intent just by the perception results?