Update FAILURE_CASES.md

Author: hkchengrex

docs/FAILURE_CASES.md

@@ -6,8 +6,12 @@ Like all methods, XMem can fail. Here, we try to show some illustrative and fran
 
 The first one is fast motion with similarly-looking objects that do not provide sufficient appearance clues for XMem to track. Below is an example from the YouTubeVOS validation set (0e8a6b63bb):
 
+https://user-images.githubusercontent.com/7107196/179459162-80b65a6c-439d-4239-819f-68804d9412e9.mp4
+
 And the source video:
 
+https://user-images.githubusercontent.com/7107196/179459166-a48884d3-58e6-4d09-9118-7df51c5305bf.mp4
+
 Technically it can be solved by using more positional and motion clues. XMem is not sufficiently proficient at those.
 
 ## Shot changes; saliency shift
@@ -16,4 +20,6 @@ Ever wondered why I did not include the final scene of Chika Dance when the roac
 
 XMem seems to be attracted to any new salient object in the scene when the (true) target object is missing. By new I mean an object that did not appear (or had a different appearance) earlier in the video -- as XMem could not have a memory representation for that object. This happens a lot if the camera shot changes.
 
+https://user-images.githubusercontent.com/7107196/179459190-d736937a-6925-4472-b46e-dcf94e1cafc0.mp4
+
 Note that the first shot change is not as problematic.