of

MERIDIAN/ONC Workshop — Victoria, BC — November 21, 2019

Detecting endangered baleen whales within acoustic recordings using R-CNNs, and how I got here...

Mark Thomas

Dalhousie University & JASCO Applied Sciences

Some background...

In late 2017/early 2018 I started working for JASCO on a detection and classification algorithm.





MammalNet:
  • no detection involved → classify vocalizations to different species using pre-extracted contours
  • eleven classes: beluga, blue, bowhead, fin, humpback, minke, right, and sei whales + walruses & bearded seals
  • shallow ConvNet developed in Tensorflow: 3 convolutional layers + 2 fully connected layers

😴  OK results, but can we cut out the contour detector?

... some (continued) background

MammalNet v2: more net, fewer mammals
  • trained on spectrograms instead of extracted contours
  • to do more we need to turn it down from 11
  • fewer species: blue, fin, and sei whales + non-biological noise & ambient noise examples
  • deeper nets: VGG-19 and ResNet-50 implemented in PyTorch
  • detection can be accomplished by sliding the classifier over a sequence of spectrograms
  • great performance! turns out: reducing your # of classes will have this affect

🤔  is this really doing "detection"?

Simonyan et al. "Very deep convolutional networks for large-scale image recognition." ICLR 2015
He et al. "Deep residual learning for image recognition." CVPR 2016

MammalNet v3: triple the v's, triple the nets

  • there have been huge advancements to the SOTA in object detection and semantic/instance segmentation:
  • in the frequency domain, marine mammal vocalization detection is more or less the same task!
  • for v2 we had full bounding box annotations, but we only used the start/end times
  • the underlying (neural net) architecture of v3 is a region-based ConvNet

screenshots taken from arXiv.org
Redmon et al. "You only look once: Unified, real-time object detection." CVPR 2016
He et al. "Mask r-cnn." ICCV 2017
Ren et al. "Faster r-cnn: Towards real-time object detection with region proposal networks." NeurIPS 2015

Region-based ConvNets a.k.a. citing Facebook four times

Girshick et al. "Rich feature hierarchies for accurate object detection and semantic segmentation." CVPR 2014
Girshick "Fast r-cnn." ICCV 2015

Mask R-CNN

 
1. the feature extraction net
  • a.k.a the "backbone"
  • used for extracting meaningful features from your input
  • users choice: ResNet, ResNeXt, VGG, MobileNet...
  • we use: ResNet-50 and then a feature pyramid network

 

2. the region proposal net
  • a second learned net
  • as the name suggests: proposes regions of interest (RoIs) directly from the feature maps of the backbone
  • RoIAlign: pooling method w/o quantization s.t. RoIs have the same size

 

3. the head net
  • a third learned net
  • provided the RoIs, computes a multi-task loss
  • classification score (softmax loss) and bounding box regression (L1-loss)

... Mask R-CNN



Acoustic recordings

  • very large collection of recordings off the coast of Atlantic Canada (Scotian Shelf)
  • 20 stations in total that were sampled at 8kHz and 250kHz
  • deployments took place in the summer/fall of 2015 and 2016
  • average deployment duration was ≈1 year
  • recorder depths ranged from 40 to 2000 meters
  • ≈19k human annotations
species label training validation test
blue whales BW 2369 502 487
fin whales FW 9554 2067 2099
sei whales SW 1210 257 227

Training details

A few training details...

Experimental results



species label instances APIoU=0.5 mAP[.5:.95] ARIoU=0.5 mAR[.5:.95]
overall - 2813 82.1 41.8 91.9 54.8
blue whales BW 487 85.7 52.8 96.2 70.9
fin whales FW 2099 75.3 30.8 89.9 40.0
sei whales SW 227 85.4 41.9 89.7 49.4

* metrics are as described in the COCO Detection Challenge

... experimental results

... experimental results

... experimental results

... experimental results

R-CNN detections for a recent recording from the Gaspé recorded in September 2019

Future work I'm doing now

Thanks × 253 to my colleagues at JASCO and Dalhousie

stills taken from "The Life Aquatic with Steve Zissou". Wes Anderson. Buena Vista Pictures, 2004

Thank you for listening!