Selective Kernel Convolution Module As described in Selective Kernel Networks (https://arxiv.org/abs/1903.06586) with some modifications.

Merge TTA predictions if applicable.

simple helper function wrapping around COCO's Python API :params: gts iterable of numpy boxes for the ground truth :params: detections itera

Create augmentor by using RandomSpatialAugmentorGenX.

Create augmentor by using RandomSpatialAugmentorGenX.

Fix the bbox that are partially or completely outside the frame. See https://github.com/prophesee-ai/prophesee-automotive-dataset-toolbox/issues/1

Transfers the fields present in dtype from an old datastructure to a new datastructure xyp should be passed as a tuple args : - d

Evaluate bbox belonging to the same class.

Filter bbox that are super wide without covering objects.

Computes a list of index ranges based on the input array of indices and a maximum length. The index ranges are computed such that the differe

Get k% pixels with the highest scores.

Get the GT and pseudo labels and their indices.

Return the bbox label of the frame at the end of this event repr. If not an exact match, return None.

Get frames (timesteps) we need to predict bbox.

Collect the IoUs and scores for this one class.

Use it only in the iter function of this class! Reason: randomized shuffling must happen within each worker. Otherwise, the same random order

Use it only in the iter function of this class!

Apply horizontal flip to bboxes.

Set cls_idx to `self.ignore_label` for low conf bbox.

match ground truth boxes and ground truth detections at all timestamps using a specified tolerance return a list of boxes vectors

May need to load images from different labeled data.

Run detection head on backbone features and do post-processing.

Predict bbox on a batch of event sequences.

Debug purpose only.

Need to find frames with labels.

Compute bbox_clamp aware velocity.

Potentially sub-sample the labeled frames.

Convert to BBOX_DTYPE. Compute `objframe_idx_2_repr/label_idx`.

utilitary function producing our data in a COCO usable format

We might track in both directions, and take and/or.

Update labels without any TTA.

Append new GT labels to self.

Update labels that undergo horizontal flip TTA.

Update labels that undergo time-flip and hflip TTA.

Set zoom_out state as zeros.

Aggregate predictions and postprocess them.

Get the datapipes that this worker should process.

For each tracker_bbox, find its closest detection_bbox in a greedy manner. Returns 3 lists of matches, unmatched_trackers, and unmatched_det

Build a dataset similar to torch.utils.data.Dataset. Each training seq contains a label.

NOTE: shape arg should include spatial dim only

Build a dataset similar to torch.utils.data.Dataset. Each training seq contains a label. Note: test dataset is always created with full sequen

Clamp bbox to image boundaries.

Make a `B`-len list of AugmentationState objects into a single dict. Returns: { 'h_flip': {'active': `B`-len list of bool},

put text in the image img

DropBlock. See https://arxiv.org/pdf/1810.12890.pdf DropBlock with an experimental gaussian noise option. This layer has been tested on a few tr

DropBlock. See https://arxiv.org/pdf/1810.12890.pdf DropBlock with an experimental gaussian noise option. Simplied from above without concern fo

Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks). This is the same as the DropConnect impl I created for E

Run the model on one event sequence and visualize it.

Compute detection KPIs on list of boxes in the numpy format, using the COCO python API https://github.com/cocodataset/cocoapi KPIs are on

Ignore the polarity of events.