- 3 x ResNet-152 (different insight resolutions and data augmentation techniques)
- 1 x DenseNet-121
Each on the preceding design, generates 1 pair of forecast for each set of harvest created by after object discovery items:
- 1 x YOLO
- 3 x Faster R-CNN with ResNet-101 as base network (different instruction iterations)
- 4 x Faster R-CNN with VGG-16 as base network (different education iterations)
- 1 x quicker R-CNN with VGG-16 as base network (trained with multi-class labeling)
As talked about above, considering that the ship credentials was correlated together with the seafood kind, it might be good for include vessel records on unit. age. harvest down rubbish), the watercraft suggestions could work as a a€?priora€? to lessen the possibilities. Ergo I utilized the soon after secret. For some image, If object alarm had not been thus confident about the forecast (for example. get back the lowest objectness rating), I joined the crop prediction with all the full picture prediction by weighted averaging. In reality, this trick is effective both on our very own validation set and general public test ready. But this secret became disastrous for all the 2nd stage exclusive examination facts which is composed of unseen and also various watercraft. This is one of several costliest wager I made regrettably they gone the wrong movement.
I made a decision to clip forecasts to a reduced certain of 0.01. It was in order to prevent big punishment of confident but incorrect responses provided by the logloss metric (i.e. -log(0) -> infinity). On top of that, We put a greater clipping continual of 0.05 for all the a€?BETa€? and a€?YFTa€? classes for cases where the forecast of a€?ALBa€? 
are large (in other words. > 0.9), since a€?ALBa€?, a€?BETa€?, and a€?YFTa€? sessions have become comparable with quite a few examples highly indistinguishable in one another.
Additional Strategies
There are several other strategies I have attempted but did not work effectively adequate to be utilized into the final remedy.
It really is rewarding to briefly discuss them here
FPN is an unique buildings which was lately launched by fb for object discovery. It is specifically made to discover items at different machines through avoid connections to combine l design drawing looks like myths. The concept is really similar to that of SSD. Making use of FPN with ResNet-101 as base circle for quicker R-CNN achieves the present state-of-the-art solitary unit listings regarding COCO bencherate of 5 fps, basically sufficient for usage in many functional solutions.
We created my own implementation of FPN with ResNet-101 in Keras and plugged it into SSD. Since FPN uses skip associations to mix feature maps at various machines, it should establish higher quality forecasts than that SSD which does not power skip connections. We anticipated that FPN is best at discovering fishes at severe machines (either exceedingly big or small). But even though the unit for some reason was able to converge, they didn’t work as well not surprisingly. I might need certainly to let it rest for additional investigation.
There clearly was another close Kaggle opposition on classifying whale species where in actuality the winners followed a novel strategy to rotate and align one’s body with the whale in order for their own mind constantly point out the exact same path, producing a a€?passporta€? image. This trick worked very well for the competitors. Which makes feel since convolutional neural system try rotational variant (well, pooling might alleviate the difficulty a bit), aligning the item interesting into the same orientation should augment category accuracy.
I e trick for the challenge. Individuals have annotated the pinnacle and tail place for each picture and posted the annotation from inside the discussion board (thank you!). In the beginning, I made use of the annotation to teach a VGG-16 regressor that immediately predicts head and tail spots through the full graphics. Of course it unsuccessful miserably. I then trained another VGG-16 regressor to predicts head-and-tail spots from cropped files, plus it worked extremely well. Our very own regressor can forecasts the actual head-and-tail roles virtually completely, as shown by the a€?red dotsa€? from the artwork lower!
