Negative Deterministic Information based Multiple Instance Learning for Weakly Supervised Object Detection and Segmentation

Weakly  supervised  object  detection  and  semanticsegmentation  with  image-level  annotations  have  attracted  ex-tensive   attention   due   to   their  high   label   efficiency.   Multipleinstance  learning  (MIL)  offers  a  feasible  solution  forthe  twotasks by treating each image as a bag with a series of instances(object  regions  or  pixels)  and  identifying  foreground  instancesthat contribute to bag classification. However, conventional MILparadigms  often  suffer  from  issues,  e.g.,  discriminative  instancedomination  and  missing  instances.In  this  paper,  weobservethat  negative  instances  usually  contain  valuable  deterministicinformation, which is the key to solving the two issues. Motivatedby  this,  we  proposea  novel  MIL  paradigm  based  on  negativedeterministic   information   (NDI),   termed   NDI-MIL,   whichisbased  on  two  core  designs  with  a  progressive  relation:  NDIcollection  and  negative  contrastive  learning.  In  NDI  collection,we  identify  and  distill  NDI  from  negative  instances  online  bya  dynamic  feature  bank.  The  collected  NDI  is  then  utilized  ina  negative  contrastive  learning  mechanism  to  locate  and  punishthose discriminative regions, by which the discriminative instancedomination and missing instances issues are effectively addressed,leading to improved object- and pixel-level localization accuracyand completeness. In addition, we design an NDI-guided instanceselection strategy to further enhance the systematic performance.Experimental  results  on  several  public  benchmarks,  includingPASCAL VOC 2007, PASCAL VOC 2012, and MS COCO, showthat  our  method  achieves  satisfactory  performance.  The  code  isavailable at: https://github.com/GC-WSL/NDI.