nizamphoenix / detr Goto Github PK

Detection Transformer leverages the transformer network(both encoder and the decoder) for Detecting Objects in Images . Facebook's researchers argue that for object detection one part of the image should be in contact with the other part of the image for greater result especially with ocluded objects and partially visible objects, and what's better than to use transformer for it.

The main ingredients of the new framework, called DEtection TRansformer or DETR,are a set-based global loss that forces unique predictions via bipartite matching, and a transformer encoder-decoder architecture. Both model and the criterion are trained

The main motive behind DETR is effectively removing the need for many hand-designed components like a non-maximum suppression procedure or anchor generation that explicitly encode prior knowledge about the task and makes the process complex and computationally expensive

DETR uses a special loss called Bipartite Matching loss where it assigns one ground truth bbox to a predicted box using a matcher , thus when fine tuning the hungarian matcher(provided by sourcecode) and also the fucntion SetCriterion which gives Bipartite matching loss for backpropogation are used.DETR calcuates three individual losses :

• Classification Loss for labels(its weight can be set by loss_ce)
• Bbox Loss (its weight can be set by loss_bbox)
• Loss for Background class

For this exercise, I used a ResNet backbone along with 5-fold stratified CV strategy to obtain results. WBF is used to ensemble the 5 models trained on 5 folds of data.