customized_models.md

Implement and Evaluate Your Own Model

To implement your model, you should inherit BaseModel defined in models/base_model.py. Here's an example:

class MyModel(BaseModel):
    @staticmethod
    def add_arguments(parser):
        # add model-specific arguments here
        ...
    
    def __init__(self, *args):
        # initialize your model here
        ...
        
    def forward(self, vic_frame, filt, *args):
        # implement the forwarding function here
        ...

Then, you should name your model and add it to SUPPORTED_MODELS in models/__init__.py. In this way, you can simply run your model by specifying --model argument in eval.py.

You can also implement your own training and evaluating framework. However, we recommend using our Evaluator() in v2x_utils/eval_utils.py and Channel() in models/model_utils/channel.py to measure the average precision and average transmission cost of your model in vehicle-infrastructure cooperative detection.

• Evaluator: Construct a evaluator by evaluator=Evaluator(pred_cls), where pred_cls is a list of the prediction classes you want to evaluate on.Then, make sure the format of model prediction is the same as ground truth labels ,which is {"boxes_3d": ..., "labels_3d": ..., "scores_3d": ...}. Then, call evaluator.add_frame(pred, label)to add the results of a single frame. When finished with all the frames, simply call evaluator.print_ap("3d"/"bev") to get the average percision results under 3d/bev view.
• Channel: Construct a channel in your model by self.channel=Channel(). Then, every time you want to transfer information from infrastructure to vehicle in your model, call self.channel.send(key, value) and value=self.channel.receive(key). Remember to use self.channel.flush() to clear all cached data in the channel after processing a single frame (otherwise the AB results will be incorrect). Use self.channel.average_bytes() to get the average transmission cost.