kg-fish

Code for Kaggle competition The Nature Conservancy Fisheries Monitoring

Goal of this competition is to detect which species of fish appears on a fishing boat, based on images captured from boat cameras of various angles. Eight target categories are available in this dataset: Albacore tuna, Bigeye tuna, Yellowfin tuna, Mahi Mahi, Opah, Sharks, Other, and No Fish. Each image has only one fish category, except that there are sometimes very small fish in the pictures that are used as bait.

Description

DeepMask/SharpMask modified from https://github.com/facebookresearch/deepmask MultiPathNet modified from https://github.com/facebookresearch/multipathnet

Requirements

• Linux (Ubuntu 16.04 LTS x86_64)
• NVidia GPU with compute capability 3.5+ and memory 4Gb+
• Python 2.7
• Python libraries
• CUDA and cuDNN v2 (or newer)
• Torch with packages: COCO API, image, tds, cjson, nnx, optim, inn, cutorch, cunn, cudnn

Training models

Data

Warning: data contains graphic contents that some may find disturbing.

Unpack data to data/raw directory. For annotations see data/dataset/annotations/ and image_labeler. If you need to label new data it must be in MS COCO format. Directory image_labeler contains code for labeling data and making dataset in Microsoft COCO format. Labeling code provided as is. *.segm.txt file row format: x1,x2,y1,y2,...,xn,yn *.lbl.txt file row format: class

Train DeepMask

cd ~/kg-fish/multipathnet/deepmask/
th train.lua -nthreads 1 -batch 8 -seed 42 -gpu 1
cp exps/deepmask/exp,batch=8,nthreads=1,seed=42/bestmodel.t7 ../../data/models/deepmask/model.t7

Train SharpMask

th train.lua -nthreads 1 -batch 4 -seed 42 -dm ../../data/models/deepmask/model.t7 -gpu 1

Compute mask proposals

Mask proposals based on DeepMask

th createProposals.lua -split train -startAt 1 -endAt 1240 -seed 42 -gpu 1 -np 1000 -dm -savedir ../../data/proposals/deepmask ../../data/models/deepmask
th createProposals.lua -split val -startAt 1 -endAt 653 -seed 42 -gpu 1 -np 1000 -dm -savedir ../../data/proposals/deepmask ../../data/models/deepmask

or Mask proposals based on SharpMask

th createProposals.lua -split train -startAt 1 -endAt 1240 -seed 42 -gpu 1 -np 1000 -savedir ../../data/proposals/deepmask ../../data/models/deepmask
th createProposals.lua -split val -startAt 1 -endAt 653 -seed 42 -gpu 1 -np 1000 -savedir ../../data/proposals/deepmask ../../data/models/deepmask

Train Multipathnet

cd ~/kg-fish/multipathnet/
cp -r ../data/dataset/* ./data
cp ../data/proposals/deepmask/* ./data/proposals/coco/deepmask

Copy model (see https://github.com/facebook/fb.resnet.torch/tree/master/pretrained )

wget https://d2j0dndfm35trm.cloudfront.net/resnet-18.t7 -O ./data/models/resnet/resnet-18.t7

To train MultiPathNet

./scripts/fish_train.sh

Test

th findFish.lua -seed 42 -dm -gpu 1 \
                -sharpmask_path ../data/models/deepmask/model.t7 \
                -multipath_path ../data/models/multipathnet/model_final.t7 \
                -imgdir ../data/raw/test_stg1 \
                -resfile ../data/submission/stg1.txt \
                -saveimg -savedir ../../result